{"id":98123,"date":"2026-09-30T20:39:06","date_gmt":"2026-09-30T17:39:06","guid":{"rendered":"https:\/\/twodots.gr\/?p=98123"},"modified":"2026-09-30T20:39:07","modified_gmt":"2026-09-30T17:39:07","slug":"quantization-aware-healing-4bit-bf16-checkpoint","status":"publish","type":"post","link":"https:\/\/twodots.gr\/bg\/quantization-aware-healing-4bit-bf16-checkpoint\/","title":{"rendered":"Quantization-Aware Healing: \u03c0\u03ce\u03c2 \u03ad\u03bd\u03b1 4-bit \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf BF16 checkpoint \u03c4\u03bf\u03c5"},"content":{"rendered":"<div class=\"td-article-lede\">\n<p><strong>\u0391\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03c0\u03c1\u03ce\u03c4\u03b1:<\/strong> \u03c4\u03bf Quantization-Aware Healing (QAH) \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1 \u03c3\u03c5\u03bc\u03c0\u03b9\u03b5\u03c3\u03bc\u03ad\u03bd\u03bf 4-bit checkpoint \u03bc\u03c0\u03bf\u03c1\u03b5\u03af, \u03bc\u03b5\u03c4\u03ac \u03b1\u03c0\u03cc \u03c3\u03c4\u03bf\u03c7\u03b5\u03c5\u03bc\u03ad\u03bd\u03b7 \u03b1\u03c0\u03cc\u03c3\u03c4\u03b1\u03be\u03b7 \u03b3\u03bd\u03ce\u03c3\u03b7\u03c2, \u03bd\u03b1 \u03be\u03b5\u03c0\u03b5\u03c1\u03ac\u03c3\u03b5\u03b9 \u03c4\u03bf BF16 checkpoint \u03b1\u03c0\u03cc \u03c4\u03bf \u03bf\u03c0\u03bf\u03af\u03bf \u03c0\u03c1\u03bf\u03ae\u03bb\u03b8\u03b5. \u03a3\u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae GPT-OSS 120B \u2192 60B \u2192 MXFP4, \u03c4\u03bf QAH \u03b9\u03c3\u03bf\u03c6\u03ac\u03c1\u03b9\u03c3\u03b5 \u03ae \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b5 \u03c4\u03bf 60B BF16 \u03c3\u03b5 7 \u03b1\u03c0\u03cc 9 benchmarks. \u0394\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03cc\u03bc\u03c9\u03c2 \u03cc\u03c4\u03b9 \u03c4\u03b1 4 bits \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03b3\u03b3\u03b5\u03bd\u03ce\u03c2 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b1: \u03c4\u03bf 4-bit \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ad\u03bb\u03b1\u03b2\u03b5 \u03ad\u03bd\u03b1 \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c4\u03bf distillation pass \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03ad\u03bb\u03b1\u03b2\u03b5 \u03c4\u03bf BF16 control.<\/p>\n<\/div>\n<p>\u0397 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03b5\u03bd\u03cc\u03c2 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf\u03c5 \u03b3\u03bb\u03c9\u03c3\u03c3\u03b9\u03ba\u03bf\u03cd \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5 \u03c3\u03c5\u03bd\u03ae\u03b8\u03c9\u03c2 \u03c0\u03b1\u03c1\u03bf\u03c5\u03c3\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c9\u03c2 \u03b1\u03bd\u03c4\u03b1\u03bb\u03bb\u03b1\u03b3\u03ae: \u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03ba\u03b1\u03b9 compute \u03bc\u03b5 \u03ba\u03ac\u03c0\u03bf\u03b9\u03bf \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c3\u03c4\u03b7\u03bd \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1. \u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03c4\u03b7\u03c2 Multiverse Computing \u03b3\u03b9\u03b1 \u03c4\u03bf <strong>Quantization-Aware Healing<\/strong> \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03b5\u03b9 \u03b1\u03bd \u03c4\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf quantization \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b3\u03af\u03bd\u03b5\u03b9 \u03c4\u03b1\u03c5\u03c4\u03cc\u03c7\u03c1\u03bf\u03bd\u03b1 \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf \u03b1\u03bd\u03ac\u03ba\u03c4\u03b7\u03c3\u03b7\u03c2 \u03b9\u03ba\u03b1\u03bd\u03bf\u03c4\u03ae\u03c4\u03c9\u03bd.<\/p>\n<p>\u03a4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b5\u03bd\u03b4\u03b9\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 \u03bf\u03bc\u03ac\u03b4\u03b5\u03c2 \u03c0\u03bf\u03c5 \u03c3\u03c7\u03b5\u03b4\u03b9\u03ac\u03b6\u03bf\u03c5\u03bd AI \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03b1, self-hosted \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ad\u03c2 \u03ba\u03b1\u03b9 \u03b1\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2: \u03c4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc 60B MXFP4 checkpoint \u03b1\u03bd\u03b1\u03c6\u03ad\u03c1\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03c6\u03bf\u03c1\u03ad\u03c2 \u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 weights \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd 60B BF16 student, \u03b5\u03bd\u03ce \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03b9\u03c3\u03c7\u03c5\u03c1\u03ae \u03b5\u03c0\u03af\u03b4\u03bf\u03c3\u03b7 \u03c3\u03b5 reasoning, \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1, tool use \u03ba\u03b1\u03b9 long context. \u0397 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03b1\u03bd\u03ac\u03b3\u03bd\u03c9\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u00ab\u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf, \u03ac\u03c1\u03b1 \u03c3\u03af\u03b3\u03bf\u03c5\u03c1\u03b1 \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf bill\u00bb, \u03b1\u03bb\u03bb\u03ac \u00ab\u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf deployment artifact \u03c0\u03bf\u03c5 \u03b1\u03be\u03af\u03b6\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03bf\u03ba\u03b9\u03bc\u03b1\u03c3\u03c4\u03b5\u03af \u03bc\u03b5 matched controls \u03ba\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc workload\u00bb.<\/p>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">\u0421\u044a\u0434\u044a\u0440\u0436\u0430\u043d\u0438\u0435<\/div>\n<ul>\n<li><a href=\"#ti-apodeiknyei\">\u03a4\u03b9 \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u2014 \u03ba\u03b1\u03b9 \u03c4\u03b9 \u03cc\u03c7\u03b9 \u2014 \u03c4\u03bf QAH<\/a><\/li>\n<li><a href=\"#sympiesi-quantization\">\u0393\u03b9\u03b1\u03c4\u03af \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03ba\u03b1\u03b9 quantization \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 healing<\/a><\/li>\n<li><a href=\"#pos-leitourgei\">\u03a0\u03ce\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03c4\u03bf Quantization-Aware Healing<\/a><\/li>\n<li><a href=\"#long-context\">\u03a0\u03ce\u03c2 \u03c7\u03c9\u03c1\u03ac \u03b7 \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03c3\u03b5 context 32k tokens<\/a><\/li>\n<li><a href=\"#ennea-benchmarks\">\u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03c4\u03b1 \u03b5\u03bd\u03bd\u03ad\u03b1 benchmarks<\/a><\/li>\n<li><a href=\"#qah-vs-qat\">QAH \u03b5\u03bd\u03b1\u03bd\u03c4\u03af\u03bf\u03bd QAT: \u03c4\u03b1\u03c7\u03cd\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc\u03c4\u03b7\u03c4\u03b1<\/a><\/li>\n<li><a href=\"#peiramatiko-keno\">\u03a4\u03bf \u03c0\u03b5\u03b9\u03c1\u03b1\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03ba\u03b5\u03bd\u03cc \u03c0\u03bf\u03c5 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03bf headline<\/a><\/li>\n<li><a href=\"#mnimi-kostos\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03bf\u03c5\u03bd \u03bc\u03bd\u03ae\u03bc\u03b7, compute \u03ba\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2<\/a><\/li>\n<li><a href=\"#production-evaluation\">\u03a0\u03ce\u03c2 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03ad\u03bd\u03b1 4-bit \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c3\u03b5 production<\/a><\/li>\n<li><a href=\"#epta-vimata\">\u03a0\u03b9\u03bb\u03bf\u03c4\u03b9\u03ba\u03ae \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae \u03c3\u03b5 \u03b5\u03c0\u03c4\u03ac \u03b2\u03ae\u03bc\u03b1\u03c4\u03b1<\/a><\/li>\n<li><a href=\"#symperasma\">\u0397 \u03bf\u03c5\u03c3\u03af\u03b1 \u03b3\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7<\/a><\/li>\n<\/ul>\n<\/div>\n<h2 id=\"ti-apodeiknyei\">\u03a4\u03b9 \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u2014 \u03ba\u03b1\u03b9 \u03c4\u03b9 \u03cc\u03c7\u03b9 \u2014 \u03c4\u03bf QAH<\/h2>\n<p>\u0397 \u03c3\u03c4\u03b5\u03bd\u03ae, \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03b9\u03c9\u03bc\u03ad\u03bd\u03b7 \u03b4\u03b9\u03b1\u03c0\u03af\u03c3\u03c4\u03c9\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7: \u03ad\u03bd\u03b1 GPT-OSS 120B \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03c3\u03c4\u03b7\u03ba\u03b5 \u03c3\u03b5 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae 60B, \u03b1\u03bd\u03b1\u03ba\u03c4\u03ae\u03b8\u03b7\u03ba\u03b5 \u03c9\u03c2 60B BF16 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b7 \u03c3\u03c5\u03bd\u03ad\u03c7\u03b5\u03b9\u03b1 \u03ba\u03b2\u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03b7\u03ba\u03b5 \u03c3\u03b5 MXFP4 \u03bc\u03b5 \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c4\u03b7 \u03b1\u03c0\u03cc\u03c3\u03c4\u03b1\u03be\u03b7 \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd \u03b1\u03c1\u03c7\u03b9\u03ba\u03cc 120B teacher. \u03a4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc 4-bit checkpoint \u03b2\u03b1\u03b8\u03bc\u03bf\u03bb\u03bf\u03b3\u03ae\u03b8\u03b7\u03ba\u03b5 \u03af\u03c3\u03bf \u03ae \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf 60B BF16 checkpoint \u03c3\u03b5 7 \u03b1\u03c0\u03cc 9 benchmarks.<\/p>\n<p>\u0397 \u03b4\u03b9\u03b1\u03c4\u03cd\u03c0\u03c9\u03c3\u03b7 \u00ab\u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf full-precision \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u00bb \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b1\u03ba\u03c1\u03af\u03b2\u03b5\u03b9\u03b1. \u03a4\u03bf \u03c3\u03b7\u03bc\u03b5\u03af\u03bf \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf <strong>\u03c3\u03c5\u03bc\u03c0\u03b9\u03b5\u03c3\u03bc\u03ad\u03bd\u03bf 60B BF16 checkpoint<\/strong>, \u03cc\u03c7\u03b9 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ac \u03c4\u03bf \u03b1\u03c1\u03c7\u03b9\u03ba\u03cc 120B. \u0395\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd, \u03c4\u03bf \u03b5\u03c0\u03af\u03c3\u03b7\u03bc\u03bf GPT-OSS 120B \u03ad\u03c7\u03b5\u03b9 \u03ae\u03b4\u03b7 \u03c4\u03b7 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b7 \u03c0\u03bb\u03b5\u03b9\u03bf\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03c9\u03bd MoE weights \u03c3\u03b5 MXFP4. \u039f teacher \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03bb\u03ae\u03c1\u03bf\u03c5\u03c2 \u03bc\u03b5\u03b3\u03ad\u03b8\u03bf\u03c5\u03c2 \u03ba\u03b1\u03b9 \u03b1\u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03c4\u03bf\u03c2, \u03cc\u03c7\u03b9 \u03ad\u03bd\u03b1 \u03c3\u03c5\u03bc\u03b2\u03b1\u03c4\u03b9\u03ba\u03cc BF16 checkpoint \u03b3\u03b9\u03b1 \u03cc\u03bb\u03b1 \u03c4\u03b1 weights.<\/p>\n<aside class=\"td-article-note\"><strong>\u0397 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ae\u03c2 \u03b1\u03bd\u03ac\u03b3\u03bd\u03c9\u03c3\u03b7:<\/strong> \u03c4\u03bf QAH \u03b5\u03af\u03bd\u03b1\u03b9 applied deployment recipe \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b1\u03c0\u03cc\u03b4\u03b5\u03b9\u03be\u03b7 \u03cc\u03c4\u03b9 \u03b7 quantization \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03b7 \u03c4\u03b7\u03c2 \u03b5\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd \u03bd\u03bf\u03b7\u03bc\u03bf\u03c3\u03cd\u03bd\u03b7. \u0394\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1 4-bit artifact \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b2\u03b3\u03b5\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf pipeline \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf BF16 \u03c3\u03b7\u03bc\u03b5\u03af\u03bf \u03b5\u03b9\u03c3\u03cc\u03b4\u03bf\u03c5 \u03c4\u03bf\u03c5 \u03cc\u03c4\u03b1\u03bd \u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c4\u03b7 teacher supervision.<\/aside>\n<p>\u0391\u03c5\u03c4\u03ae \u03b7 \u03b4\u03b9\u03ac\u03ba\u03c1\u03b9\u03c3\u03b7 \u03ad\u03c7\u03b5\u03b9 \u03c3\u03b7\u03bc\u03b1\u03c3\u03af\u03b1 \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03c4\u03bf governance \u03c4\u03c9\u03bd open-weight \u03c3\u03c5\u03c3\u03c4\u03b7\u03bc\u03ac\u03c4\u03c9\u03bd. \u038c\u03c0\u03c9\u03c2 \u03b5\u03be\u03b7\u03b3\u03b5\u03af \u03c4\u03bf \u03ac\u03c1\u03b8\u03c1\u03bf \u03b3\u03b9\u03b1 <a href=\"https:\/\/twodots.gr\/model-cards-open-weight-ai-governance\/\">\u03c4\u03b1 \u03cc\u03c1\u03b9\u03b1 \u03c4\u03c9\u03bd model cards \u03c3\u03c4\u03b1 open-weight AI \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1<\/a>, \u03b7 \u03bf\u03bd\u03bf\u03bc\u03b1\u03c3\u03af\u03b1 \u03c4\u03bf\u03c5 checkpoint \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af. \u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ae\u03c2 lineage, training recipe, evaluation protocol \u03ba\u03b1\u03b9 \u03c3\u03b1\u03c6\u03ae \u03cc\u03c1\u03b9\u03b1 \u03c3\u03c4\u03bf \u03c4\u03b9 \u03c5\u03c0\u03bf\u03c3\u03c4\u03b7\u03c1\u03af\u03b6\u03bf\u03c5\u03bd \u03bf\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2.<\/p>\n<h2 id=\"sympiesi-quantization\">\u0393\u03b9\u03b1\u03c4\u03af \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03ba\u03b1\u03b9 quantization \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 healing<\/h2>\n<p>\u0397 \u03b4\u03bf\u03bc\u03b9\u03ba\u03ae \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03bf \u03bc\u03ad\u03b3\u03b5\u03b8\u03bf\u03c2 \u03bc\u03b9\u03b1\u03c2 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae\u03c2 \u03b1\u03c6\u03b1\u03b9\u03c1\u03ce\u03bd\u03c4\u03b1\u03c2 \u03ae \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03c2 layers, attention heads, neurons \u03ae \u03ac\u03bb\u03bb\u03b5\u03c2 \u03b4\u03b9\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03b9\u03c2. \u0397 quantization \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c1\u03b9\u03b8\u03bc\u03b7\u03c4\u03b9\u03ba\u03ae \u03b1\u03ba\u03c1\u03af\u03b2\u03b5\u03b9\u03b1 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03bf\u03c0\u03bf\u03af\u03b1 \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03cd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03c4\u03b1 weights. \u039f\u03b9 \u03b4\u03cd\u03bf \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ad\u03c2 \u03b5\u03be\u03bf\u03b9\u03ba\u03bf\u03bd\u03bf\u03bc\u03bf\u03cd\u03bd \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03bf\u03cd\u03c2 \u03c0\u03cc\u03c1\u03bf\u03c5\u03c2, \u03b1\u03bb\u03bb\u03ac \u03cc\u03c4\u03b1\u03bd \u03b5\u03c6\u03b1\u03c1\u03bc\u03cc\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b4\u03b9\u03b1\u03b4\u03bf\u03c7\u03b9\u03ba\u03ac \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03b5\u03c0\u03b9\u03b2\u03b1\u03c1\u03cd\u03bd\u03bf\u03c5\u03bd reasoning, \u03bc\u03b1\u03b8\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03b5\u03c0\u03af\u03bb\u03c5\u03c3\u03b7, \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1, tool use \u03ba\u03b1\u03b9 \u03ba\u03b1\u03c4\u03b1\u03bd\u03cc\u03b7\u03c3\u03b7 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf\u03c5 context.<\/p>\n<p>\u0393\u03b9\u2019 \u03b1\u03c5\u03c4\u03cc \u03ad\u03bd\u03b1 \u03c3\u03bf\u03b2\u03b1\u03c1\u03cc production pipeline \u03b4\u03b5\u03bd \u03c4\u03b5\u03bb\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03b1\u03bd\u03b1\u03b3\u03ba\u03b1\u03c3\u03c4\u03b9\u03ba\u03ac \u03bc\u03b5 \u03c4\u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7. \u03a0\u03c1\u03bf\u03c3\u03b8\u03ad\u03c4\u03b5\u03b9 recovery \u03ae healing \u03ce\u03c3\u03c4\u03b5 \u03c4\u03bf \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf checkpoint \u03bd\u03b1 \u03b1\u03bd\u03b1\u03ba\u03c4\u03ae\u03c3\u03b5\u03b9 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ad\u03c2 \u03c0\u03bf\u03c5 \u03c7\u03ac\u03b8\u03b7\u03ba\u03b1\u03bd. \u03a4\u03bf \u03b5\u03c1\u03ce\u03c4\u03b7\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bb\u03ce\u03c2 \u03b1\u03bd \u03b7 \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03ad\u03c7\u03b5\u03b9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2, \u03b1\u03bb\u03bb\u03ac \u03c0\u03bf\u03b9\u03bf\u03c2 teacher, \u03c0\u03bf\u03b9\u03b1 loss function \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1 \u03b4\u03af\u03bd\u03bf\u03c5\u03bd \u03c3\u03c4\u03bf compressed \u03ba\u03b1\u03b9 quantized \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c4\u03b7\u03bd \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03c0\u03b9\u03b8\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03bd\u03b1 \u03b5\u03c0\u03b1\u03bd\u03ad\u03bb\u03b8\u03b5\u03b9.<\/p>\n<p>\u041d\u0430 \u0430\u0434\u0440\u0435\u0441 <strong>quantization-aware training (QAT)<\/strong>, fake-quantization operators \u03bc\u03c0\u03b1\u03af\u03bd\u03bf\u03c5\u03bd \u03c3\u03c4\u03bf forward pass \u03ba\u03b1\u03b9 \u03c4\u03bf training \u03c3\u03c5\u03bd\u03b5\u03c7\u03af\u03b6\u03b5\u03b9 \u03bc\u03b5 task loss, \u03c3\u03c5\u03bd\u03ae\u03b8\u03c9\u03c2 cross-entropy. \u03a3\u03c4\u03bf <strong>quantization-aware distillation (QAD)<\/strong>, \u03bf quantized student \u03bc\u03b9\u03bc\u03b5\u03af\u03c4\u03b1\u03b9 \u03c4\u03b7\u03bd \u03ba\u03b1\u03c4\u03b1\u03bd\u03bf\u03bc\u03ae \u03b5\u03be\u03cc\u03b4\u03bf\u03c5 \u03b5\u03bd\u03cc\u03c2 frozen teacher \u03bc\u03ad\u03c3\u03c9 KL divergence. \u0397 \u03ba\u03bb\u03b1\u03c3\u03b9\u03ba\u03ae QAD \u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03c4\u03b1\u03b9\u03c1\u03b9\u03ac\u03b6\u03b5\u03b9 \u03cc\u03c4\u03b1\u03bd teacher \u03ba\u03b1\u03b9 student \u03ad\u03c7\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03c0\u03c1\u03b9\u03bd \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03ac \u03c4\u03b7\u03bd quantization.<\/p>\n<p>\u039c\u03b5 \u03b4\u03bf\u03bc\u03b9\u03ba\u03ae \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7, \u03cc\u03bc\u03c9\u03c2, \u03c4\u03bf 60B BF16 checkpoint \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b1 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03c5\u03bc\u03ad\u03bd\u03bf 60B \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf. \u0395\u03af\u03bd\u03b1\u03b9 \u03ae\u03b4\u03b7 \u03bc\u03b9\u03b1 \u03b1\u03bd\u03b1\u03ba\u03c4\u03b7\u03bc\u03ad\u03bd\u03b7 \u03c0\u03c1\u03bf\u03c3\u03ad\u03b3\u03b3\u03b9\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf\u03c5 120B. \u0391\u03bd \u03b3\u03af\u03bd\u03b5\u03b9 teacher \u03b3\u03b9\u03b1 \u03c4\u03bf \u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf, \u03bf student \u03ba\u03bb\u03b7\u03c1\u03bf\u03bd\u03bf\u03bc\u03b5\u03af \u03c4\u03bf \u03b4\u03b9\u03ba\u03cc \u03c4\u03bf\u03c5 ceiling. \u0395\u03ba\u03b5\u03af \u03b1\u03ba\u03c1\u03b9\u03b2\u03ce\u03c2 \u03c0\u03b1\u03c1\u03b5\u03bc\u03b2\u03b1\u03af\u03bd\u03b5\u03b9 \u03c4\u03bf QAH.<\/p>\n<h2 id=\"pos-leitourgei\">\u03a0\u03ce\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03c4\u03bf Quantization-Aware Healing<\/h2>\n<p>\u03a4\u03bf QAH \u03c0\u03b1\u03c1\u03b1\u03ba\u03ac\u03bc\u03c0\u03c4\u03b5\u03b9 \u03c4\u03bf \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf BF16 checkpoint \u03c9\u03c2 teacher. \u039f student \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03c3\u03c5\u03bc\u03c0\u03b9\u03b5\u03c3\u03bc\u03ad\u03bd\u03b7 60B \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03bc\u03b5 MXFP4 fake quantizers, \u03b5\u03bd\u03ce \u03bf teacher \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03c4\u03bf \u03b1\u03c1\u03c7\u03b9\u03ba\u03cc \u03b1\u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03c4\u03bf GPT-OSS 120B. \u0395\u03c0\u03b5\u03b9\u03b4\u03ae \u03b7 \u03b1\u03c0\u03cc\u03c3\u03c4\u03b1\u03be\u03b7 \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c0\u03ac\u03bd\u03c9 \u03c3\u03c4\u03b1 output logits, teacher \u03ba\u03b1\u03b9 student \u03b4\u03b5\u03bd \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03ad\u03c7\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ae \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae.<\/p>\n<p>\u039f student \u03b4\u03b5\u03bd \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03b5\u03c4\u03b1\u03b9 \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 hard labels. \u03a0\u03c1\u03bf\u03c3\u03c0\u03b1\u03b8\u03b5\u03af \u03bd\u03b1 \u03c0\u03c1\u03bf\u03c3\u03b5\u03b3\u03b3\u03af\u03c3\u03b5\u03b9 \u03c4\u03b7\u03bd \u03ba\u03b1\u03c4\u03b1\u03bd\u03bf\u03bc\u03ae \u03c0\u03b9\u03b8\u03b1\u03bd\u03bf\u03c4\u03ae\u03c4\u03c9\u03bd \u03c0\u03bf\u03c5 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03bf teacher \u03b3\u03b9\u03b1 \u03c4\u03bf \u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03bf token. \u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af offline top-100 teacher logits \u03ba\u03b1\u03b9 KL divergence, \u03ce\u03c3\u03c4\u03b5 \u03bf \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf\u03c2 teacher \u03bd\u03b1 \u03bc\u03b7 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 training step.<\/p>\n<div class=\"td-comparison td-comparison-cards td-comparison-cards--horizontal\">\n<div class=\"td-comparison-grid td-comparison-grid-three\">\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">60B BF16 source<\/p>\n<p>\u0388\u03c7\u03b5\u03b9 \u03ae\u03b4\u03b7 \u03b1\u03bd\u03b1\u03ba\u03c4\u03b7\u03b8\u03b5\u03af \u03b1\u03c0\u03cc \u03c4\u03b7 \u03b4\u03bf\u03bc\u03b9\u03ba\u03ae \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7 \u03bc\u03ad\u03c3\u03c9 distillation. \u0395\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf checkpoint \u03c0\u03bf\u03c5 \u03bc\u03c0\u03b1\u03af\u03bd\u03b5\u03b9 \u03c3\u03c4\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf quantization, \u03cc\u03c7\u03b9 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b1 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03c5\u03bc\u03ad\u03bd\u03bf 60B \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">16-bit weights<\/span><span class=\"td-badge\">Recovery pass<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card td-platform-card--navy\">\n<p class=\"td-comparison-title\">60B MXFP4 \u03bc\u03b5 QAH<\/p>\n<p>\u039b\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf distillation pass \u03b1\u03c0\u03b5\u03c5\u03b8\u03b5\u03af\u03b1\u03c2 \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd \u03b1\u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03c4\u03bf 120B teacher, \u03b5\u03bd\u03ce \u03c4\u03b1 fake quantizers \u03c0\u03c1\u03bf\u03c3\u03bf\u03bc\u03bf\u03b9\u03ce\u03bd\u03bf\u03c5\u03bd \u03c4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc 4-bit deployment.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">4-bit artifact<\/span><span class=\"td-badge\">Teacher logits<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">\u03a3\u03c9\u03c3\u03c4\u03cc production control<\/p>\n<p>\u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 BF16 arm \u03bc\u03b5 \u03af\u03b4\u03b9\u03bf teacher, data \u03ba\u03b1\u03b9 steps, \u03ba\u03b1\u03b8\u03ce\u03c2 \u03ba\u03b1\u03b9 post-training quantization \u03b1\u03c5\u03c4\u03bf\u03cd \u03c4\u03bf\u03c5 arm, \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03bc\u03bf\u03bd\u03c9\u03b8\u03b5\u03af \u03b7 \u03b1\u03be\u03af\u03b1 \u03c4\u03bf\u03c5 healing \u03c5\u03c0\u03cc quantization.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Matched compute<\/span><span class=\"td-badge\">\u038a\u03b4\u03b9\u03bf evaluation<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u0397 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03c4\u03bf\u03c5 \u03b1\u03c1\u03c7\u03b9\u03ba\u03bf\u03cd teacher \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03c4\u03bf quantization stage \u03c3\u03b5 \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03b7 \u03b5\u03c5\u03ba\u03b1\u03b9\u03c1\u03af\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac\u03c2 \u03b3\u03bd\u03ce\u03c3\u03b7\u03c2. \u03a4\u03bf MXFP4 checkpoint \u03b4\u03b5\u03bd \u03c0\u03c1\u03bf\u03c3\u03c0\u03b1\u03b8\u03b5\u03af \u03bc\u03cc\u03bd\u03bf \u03bd\u03b1 \u03b1\u03bd\u03c4\u03ad\u03be\u03b5\u03b9 \u03c4\u03bf\u03bd \u03b1\u03c1\u03b9\u03b8\u03bc\u03b7\u03c4\u03b9\u03ba\u03cc \u03b8\u03cc\u03c1\u03c5\u03b2\u03bf. \u039b\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 supervision \u03c0\u03bf\u03c5 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03c0\u03b1\u03bd\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03b7 \u03bf\u03c0\u03bf\u03af\u03b1 \u03b4\u03b5\u03bd \u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c0\u03bb\u03ae\u03c1\u03c9\u03c2 \u03c3\u03c4\u03bf\u03bd \u03c0\u03c1\u03ce\u03c4\u03bf \u03ba\u03cd\u03ba\u03bb\u03bf recovery.<\/p>\n<h2 id=\"long-context\">\u03a0\u03ce\u03c2 \u03c7\u03c9\u03c1\u03ac \u03b7 \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03c3\u03b5 context 32k tokens<\/h2>\n<p>\u0397 \u03b1\u03c0\u03cc\u03c3\u03c4\u03b1\u03be\u03b7 \u03c3\u03b5 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf context \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03b1\u03b9\u03c4\u03b7\u03c4\u03b9\u03ba\u03ae. \u0391\u03bd \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03c5\u03c4\u03b5\u03af \u03b7 \u03c0\u03bb\u03ae\u03c1\u03b7\u03c2 \u03ba\u03b1\u03c4\u03b1\u03bd\u03bf\u03bc\u03ae \u03c0\u03b9\u03b8\u03b1\u03bd\u03bf\u03c4\u03ae\u03c4\u03c9\u03bd \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 token \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c4\u03bf vocabulary, \u03c4\u03bf \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf tensor \u03ba\u03b1\u03b9 \u03c4\u03bf autograd graph \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03b5\u03be\u03b1\u03bd\u03c4\u03bb\u03ae\u03c3\u03bf\u03c5\u03bd \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03b1 \u03c4\u03b7 GPU memory. \u03a4\u03bf QAH \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c4\u03b7\u03bd offline top-k \u03ba\u03b1\u03b9 fused chunked-KL \u03c0\u03c1\u03bf\u03c3\u03ad\u03b3\u03b3\u03b9\u03c3\u03b7 \u03c4\u03b7\u03c2 \u03c3\u03c5\u03bd\u03bf\u03b4\u03b5\u03c5\u03c4\u03b9\u03ba\u03ae\u03c2 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1\u03c2 \u03c4\u03b7\u03c2 \u03af\u03b4\u03b9\u03b1\u03c2 \u03bf\u03bc\u03ac\u03b4\u03b1\u03c2.<\/p>\n<p>\u03a4\u03b1 teacher logits \u03c0\u03c1\u03bf\u03cb\u03c0\u03bf\u03bb\u03bf\u03b3\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03ba\u03c1\u03b1\u03c4\u03b9\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03b3\u03b9\u03b1 \u03c4\u03b1 \u03ba\u03bf\u03c1\u03c5\u03c6\u03b1\u03af\u03b1 tokens, \u03b5\u03bd\u03ce \u03c4\u03bf KL loss \u03ba\u03b1\u03b9 \u03bf\u03b9 gradients \u03c3\u03c5\u03c3\u03c3\u03c9\u03c1\u03b5\u03cd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03b5 chunks \u03c4\u03b7\u03c2 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03af\u03b1\u03c2. \u0391\u03c5\u03c4\u03cc \u03b5\u03c0\u03b9\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 training \u03bc\u03b5 sequence length 32k \u03c3\u03c4\u03bf \u03b1\u03bd\u03b1\u03c6\u03b5\u03c1\u03cc\u03bc\u03b5\u03bd\u03bf setup. \u0397 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03ba\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf long-context healing \u03b5\u03c6\u03b9\u03ba\u03c4\u03cc\u00b7 \u03b4\u03b5\u03bd \u03b5\u03b3\u03b3\u03c5\u03ac\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03b7 \u03c4\u03b7\u03c2 \u03cc\u03c4\u03b9 \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03b8\u03b1 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c3\u03c9\u03c3\u03c4\u03ac \u03ba\u03ac\u03b8\u03b5 \u03b8\u03ad\u03c3\u03b7 \u03c4\u03bf\u03c5 context.<\/p>\n<p>\u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03bf\u03bc\u03ac\u03b4\u03b1, \u03c4\u03bf \u03c3\u03b7\u03bc\u03b5\u03af\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03be\u03af\u03c3\u03bf\u03c5 \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03cc \u03bc\u03b5 \u03c4\u03bf benchmark. \u0388\u03bd\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03ba\u03b1\u03bb\u03ae \u03b8\u03b5\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03ae \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ac \u03b1\u03bb\u03bb\u03ac \u03bd\u03b1 \u03bc\u03b7\u03bd \u03c7\u03c9\u03c1\u03ac \u03c3\u03c4\u03bf \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03bf hardware \u03ae \u03bd\u03b1 \u03b5\u03c0\u03b9\u03b2\u03b1\u03c1\u03cd\u03bd\u03b5\u03b9 \u03b4\u03c5\u03c3\u03b1\u03bd\u03ac\u03bb\u03bf\u03b3\u03b1 \u03c4\u03bf serving. \u0397 \u03af\u03b4\u03b9\u03b1 \u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03b7\u03bd <a href=\"https:\/\/twodots.gr\/glide-yvridiki-prosochi-meionei-kostos-llm-megalou-context\/\">\u03c5\u03b2\u03c1\u03b9\u03b4\u03b9\u03ba\u03ae \u03c0\u03c1\u03bf\u03c3\u03bf\u03c7\u03ae \u03b3\u03b9\u03b1 LLM \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf\u03c5 context<\/a>: \u03b7 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03ba\u03b1\u03b9\u03bd\u03bf\u03c4\u03bf\u03bc\u03af\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03bc\u03b1\u03b6\u03af \u03bc\u03b5 \u03c4\u03bf memory envelope \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b1\u03c0\u03bf\u03ba\u03bf\u03bc\u03bc\u03ad\u03bd\u03b1 \u03b1\u03c0\u03cc \u03b1\u03c5\u03c4\u03cc.<\/p>\n<h2 id=\"ennea-benchmarks\">\u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03c4\u03b1 \u03b5\u03bd\u03bd\u03ad\u03b1 benchmarks<\/h2>\n<p>\u0397 \u03b4\u03b7\u03bc\u03bf\u03c3\u03b9\u03b5\u03c5\u03bc\u03ad\u03bd\u03b7 \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03c0\u03b5\u03c1\u03b9\u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf\u03bd \u03b1\u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03c4\u03bf 120B teacher, \u03c4\u03bf \u03b1\u03bd\u03b1\u03ba\u03c4\u03b7\u03bc\u03ad\u03bd\u03bf 60B BF16 checkpoint \u03ba\u03b1\u03b9 \u03c4\u03bf 60B MXFP4 checkpoint \u03bc\u03b5\u03c4\u03ac \u03c4\u03bf QAH. \u03a4\u03bf \u03c4\u03b5\u03bb\u03b5\u03c5\u03c4\u03b1\u03af\u03bf \u03b9\u03c3\u03bf\u03c6\u03b1\u03c1\u03af\u03b6\u03b5\u03b9 \u03ae \u03c5\u03c0\u03b5\u03c1\u03ad\u03c7\u03b5\u03b9 \u03c4\u03bf\u03c5 BF16 \u03c3\u03b5 \u03b5\u03c0\u03c4\u03ac \u03b1\u03c0\u03cc \u03b5\u03bd\u03bd\u03ad\u03b1 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2. \u039f\u03b9 \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b5\u03c2 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ad\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c4\u03bf AA-LCR \u03b3\u03b9\u03b1 long-context reasoning, \u03bc\u03b5 42,7 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 35,3, \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf AIME 2025 \u03b3\u03b9\u03b1 \u03bc\u03b1\u03b8\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac, \u03bc\u03b5 76,3 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 70,7.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-kicker\">\u0394\u03b7\u03bc\u03bf\u03c3\u03b9\u03b5\u03c5\u03bc\u03ad\u03bd\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03c4\u03bf\u03c5 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf\u03c5 pipeline<\/p>\n<p class=\"td-chart-title\">\u03a4\u03bf QAH \u03b1\u03bd\u03b1\u03ba\u03c4\u03ac \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1, \u03b1\u03bb\u03bb\u03ac \u03bf\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03c9\u03c3\u03c4\u03cc scope<\/p>\n<p class=\"td-chart-intro\">\u039f\u03b9 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03c4\u03b7\u03c2 Multiverse Computing. \u0394\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc\u03c2 \u03b4\u03b5\u03af\u03ba\u03c4\u03b7\u03c2 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 4-bit \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf, hardware \u03ae \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc workload.<\/p>\n<\/div>\n<div class=\"td-metric-grid\">\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">7\/9<\/span><span class=\"td-metric-label\">benchmarks \u03cc\u03c0\u03bf\u03c5 \u03c4\u03bf 60B MXFP4 QAH \u03b9\u03c3\u03bf\u03c6\u03ac\u03c1\u03b9\u03c3\u03b5 \u03ae \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf 60B BF16 source<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">\u22484\u00d7<\/span><span class=\"td-metric-label\">\u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 weights \u03b3\u03b9\u03b1 \u03c4\u03bf MXFP4 checkpoint \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c4\u03bf\u03c5 BF16 student<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">100 \/ 700<\/span><span class=\"td-metric-label\">\u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 steps \u03b3\u03b9\u03b1 \u03c4\u03bf peak \u03c4\u03bf\u03c5 QAH \u03ba\u03b1\u03b9 \u03c4\u03bf\u03c5 matched QAT \u03c3\u03c4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 GPT-OSS 9B<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">32k tokens<\/span><span class=\"td-metric-label\">sequence length \u03c4\u03bf\u03c5 QAH training \u03bc\u03b5 offline top-k logits \u03ba\u03b1\u03b9 chunked KL loss<\/span><\/div>\n<\/div>\n<p class=\"td-chart-source\">\u03a0\u03b7\u03b3\u03ad\u03c2: QAH paper \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03bf\u03b4\u03b5\u03c5\u03c4\u03b9\u03ba\u03ae \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b3\u03b9\u03b1 efficient knowledge distillation.<\/p>\n<\/div>\n<p>\u03a3\u03c4\u03bf MMLU-Pro \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf SciCode \u03c4\u03bf QAH \u03bc\u03ad\u03bd\u03b5\u03b9 \u03c0\u03af\u03c3\u03c9 \u03b1\u03c0\u03cc \u03c4\u03bf BF16 \u03ba\u03b1\u03c4\u03ac 0,2 \u03ba\u03b1\u03b9 1,4 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b1. \u03a3\u03b5 \u03c3\u03c7\u03ad\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03bf\u03bd 120B teacher, \u03c4\u03bf 60B QAH \u03c6\u03c4\u03ac\u03bd\u03b5\u03b9 \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ac \u03c4\u03b7\u03bd \u03b5\u03c0\u03af\u03b4\u03bf\u03c3\u03b7 \u03c3\u03c4\u03bf LiveCodeBench, 66,5 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 66,0, \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03c0\u03bf\u03c5 \u03bf\u03b9 \u03af\u03b4\u03b9\u03bf\u03b9 \u03bf\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03b8\u03b5\u03c9\u03c1\u03bf\u03cd\u03bd \u03b5\u03bd\u03c4\u03cc\u03c2 run-to-run noise. \u03a3\u03c4\u03bf GPQA Diamond \u03c6\u03c4\u03ac\u03bd\u03b5\u03b9 67,4 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 69,0.<\/p>\n<p>\u03a4\u03b1 benchmark scores \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 SLA. \u03a4\u03bf <a href=\"https:\/\/twodots.gr\/ai-benchmark-harness-allazei-nikiti\/\">evaluation harness \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03bb\u03bb\u03ac\u03be\u03b5\u03b9 \u03c4\u03bf\u03bd \u03c6\u03b1\u03b9\u03bd\u03bf\u03bc\u03b5\u03bd\u03b9\u03ba\u03cc \u03bd\u03b9\u03ba\u03b7\u03c4\u03ae<\/a>, \u03b5\u03bd\u03ce task mix, prompts, decoding, hardware \u03ba\u03b1\u03b9 serving stack \u03b5\u03c0\u03b7\u03c1\u03b5\u03ac\u03b6\u03bf\u03c5\u03bd \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1. \u0393\u03b9\u2019 \u03b1\u03c5\u03c4\u03cc \u03bf\u03b9 \u03b5\u03bd\u03bd\u03ad\u03b1 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03ae\u03bc\u03b1 \u03b3\u03b9\u03b1 \u03c0\u03b5\u03c1\u03b1\u03b9\u03c4\u03ad\u03c1\u03c9 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7, \u03cc\u03c7\u03b9 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03b1\u03b3\u03bf\u03c1\u03ac\u03c2 \u03ae deployment.<\/p>\n<h2 id=\"qah-vs-qat\">QAH \u03b5\u03bd\u03b1\u03bd\u03c4\u03af\u03bf\u03bd QAT: \u03c4\u03b1\u03c7\u03cd\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc\u03c4\u03b7\u03c4\u03b1<\/h2>\n<p>\u0393\u03b9\u03b1 \u03bd\u03b1 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03bf\u03c5\u03bd \u03c4\u03b9\u03c2 \u03b4\u03cd\u03bf loss functions \u03c5\u03c0\u03cc matched \u03c3\u03c5\u03bd\u03b8\u03ae\u03ba\u03b5\u03c2, \u03bf\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ad\u03c2 \u03ba\u03b2\u03ac\u03bd\u03c4\u03b9\u03c3\u03b1\u03bd \u03ad\u03bd\u03b1 GPT-OSS 9B \u03c3\u03b5 MXFP4 \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u03b1\u03ba\u03bf\u03bb\u03bf\u03cd\u03b8\u03b7\u03c3\u03b1\u03bd \u03c4\u03bf\u03bd \u03bc\u03ad\u03c3\u03bf \u03cc\u03c1\u03bf \u03b5\u03c0\u03af\u03b4\u03bf\u03c3\u03b7\u03c2 \u03c3\u03c4\u03b1 MMLU-Pro, LiveCodeBench \u03ba\u03b1\u03b9 GPQA Diamond. \u03a4\u03bf QAH \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 peak 54,9 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c3\u03c4\u03b1 100 steps. \u03a4\u03bf QAT \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 \u03c3\u03b5 \u03c0\u03b1\u03c1\u03cc\u03bc\u03bf\u03b9\u03bf peak 54,6 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c3\u03c4\u03b1 700 steps.<\/p>\n<p>\u039c\u03b5\u03c4\u03ac \u03c4\u03b7\u03bd \u03ba\u03bf\u03c1\u03c5\u03c6\u03ae, \u03c4\u03bf QAH \u03c0\u03b1\u03c1\u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03b5\u03bd\u03c4\u03cc\u03c2 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03b4\u03cd\u03bf \u03bc\u03bf\u03bd\u03ac\u03b4\u03c9\u03bd \u03ad\u03c9\u03c2 \u03c4\u03b1 1.200 steps, \u03b5\u03bd\u03ce \u03c4\u03bf QAT \u03ad\u03c7\u03b1\u03c3\u03b5 \u03c3\u03c7\u03b5\u03b4\u03cc\u03bd 19 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2. \u039f\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03b1\u03c0\u03bf\u03b4\u03af\u03b4\u03bf\u03c5\u03bd \u03c4\u03b7 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03c3\u03c4\u03bf objective: \u03c4\u03bf fixed teacher distribution \u03b4\u03b5\u03bd \u03c3\u03c5\u03bd\u03b5\u03c7\u03af\u03b6\u03b5\u03b9 \u03bd\u03b1 \u03c9\u03b8\u03b5\u03af \u03c4\u03bf\u03bd student \u03c0\u03c1\u03bf\u03c2 hard labels \u03b1\u03c6\u03bf\u03cd \u03c0\u03bb\u03b7\u03c3\u03b9\u03ac\u03c3\u03b5\u03b9 \u03c4\u03bf\u03bd teacher. \u0391\u03c5\u03c4\u03cc \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b5\u03c5\u03b1\u03b9\u03c3\u03b8\u03b7\u03c3\u03af\u03b1 \u03c3\u03c4\u03bf early stopping \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1, \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03ba\u03b1\u03c4\u03b1\u03c1\u03b3\u03b5\u03af \u03c4\u03b7\u03bd \u03b1\u03bd\u03ac\u03b3\u03ba\u03b7 \u03b3\u03b9\u03b1 held-out validation.<\/p>\n<p>\u0397 \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03b1\u03c5\u03c4\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03b7 \u03b1\u03c0\u03cc \u03c4\u03bf 60B headline \u03c9\u03c2 \u03c0\u03c1\u03bf\u03c2 \u03c4\u03b7\u03bd loss function, \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae QAH \u03ba\u03b1\u03b9 QAT \u03ad\u03c7\u03bf\u03c5\u03bd matched \u03c3\u03c5\u03bd\u03b8\u03ae\u03ba\u03b5\u03c2. \u03a0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03cc\u03bc\u03c9\u03c2 \u03ad\u03bd\u03b1 \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf, \u03c4\u03c1\u03af\u03b1 benchmarks \u03ba\u03b1\u03b9 \u03ad\u03bd\u03b1 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf training setup. \u0394\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03b1\u03c0\u03b5\u03af \u03c3\u03b5 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03cc \u03b9\u03c3\u03c7\u03c5\u03c1\u03b9\u03c3\u03bc\u03cc \u03cc\u03c4\u03b9 \u03c4\u03bf QAH \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03ac\u03bd\u03c4\u03bf\u03c4\u03b5 \u03b5\u03c0\u03c4\u03ac \u03c6\u03bf\u03c1\u03ad\u03c2 \u03c6\u03b8\u03b7\u03bd\u03cc\u03c4\u03b5\u03c1\u03bf.<\/p>\n<h2 id=\"peiramatiko-keno\">\u03a4\u03bf \u03c0\u03b5\u03b9\u03c1\u03b1\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03ba\u03b5\u03bd\u03cc \u03c0\u03bf\u03c5 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03bf headline<\/h2>\n<p>\u039c\u03b5\u03c4\u03ac \u03c4\u03b7 \u03b4\u03b7\u03bc\u03bf\u03c3\u03af\u03b5\u03c5\u03c3\u03b7, \u03c3\u03c7\u03bf\u03bb\u03b9\u03b1\u03c3\u03c4\u03ae\u03c2 \u03c3\u03c4\u03bf Hugging Face \u03b5\u03c0\u03b9\u03c3\u03ae\u03bc\u03b1\u03bd\u03b5 \u03cc\u03c4\u03b9 \u03b7 \u03ba\u03b5\u03bd\u03c4\u03c1\u03b9\u03ba\u03ae 60B \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03b4\u03b5\u03bd \u03ad\u03c7\u03b5\u03b9 \u03c0\u03bb\u03ae\u03c1\u03c9\u03c2 \u03b9\u03c3\u03bf\u03b4\u03cd\u03bd\u03b1\u03bc\u03bf control. \u03a4\u03bf 60B BF16 checkpoint \u03b5\u03af\u03c7\u03b5 \u03bb\u03ac\u03b2\u03b5\u03b9 \u03ad\u03bd\u03b1 distillation pass \u03ba\u03b1\u03c4\u03ac \u03c4\u03bf recovery. \u03a4\u03bf 60B MXFP4 \u03b5\u03af\u03c7\u03b5 \u03bb\u03ac\u03b2\u03b5\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf recovery \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf pass \u03b1\u03c0\u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c3\u03c4\u03bf\u03bd 120B teacher. \u0386\u03c1\u03b1 \u03b7 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03bc\u03bf\u03bd\u03ce\u03bd\u03b5\u03b9 \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c4\u03bf\u03c5 healing \u03c5\u03c0\u03cc quantization \u03b1\u03c0\u03cc \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c4\u03bf\u03c5 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf\u03c5 training \u03bc\u03b5 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc teacher.<\/p>\n<p>\u039f \u03c3\u03c5\u03bd-\u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03ad\u03b1\u03c2 Antonio Tiene \u03b1\u03bd\u03b1\u03b3\u03bd\u03ce\u03c1\u03b9\u03c3\u03b5 \u03b4\u03b7\u03bc\u03cc\u03c3\u03b9\u03b1 \u03c4\u03bf \u03c3\u03b7\u03bc\u03b5\u03af\u03bf. \u03a4\u03bf \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03c0\u03bf\u03c5 \u03c5\u03c0\u03bf\u03c3\u03c4\u03b7\u03c1\u03af\u03b6\u03b5\u03b9 \u03b7 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c4\u03b5\u03bd\u03cc\u03c4\u03b5\u03c1\u03bf: \u03c4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc 4-bit checkpoint \u03b2\u03b1\u03b8\u03bc\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03b1 \u03b1\u03c0\u03cc \u03c4\u03bf BF16 checkpoint \u03c0\u03bf\u03c5 \u03bc\u03c0\u03ae\u03ba\u03b5 \u03c3\u03c4\u03bf quantization stage. \u0393\u03b9\u03b1 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03b7 \u03b1\u03b9\u03c4\u03b9\u03ce\u03b4\u03b7 \u03b1\u03c0\u03cc\u03b4\u03b5\u03b9\u03be\u03b7 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 BF16 \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf pass \u03bc\u03b5 \u03af\u03b4\u03b9\u03bf teacher, data \u03ba\u03b1\u03b9 steps, \u03ba\u03b1\u03b8\u03ce\u03c2 \u03ba\u03b1\u03b9 post-training quantization \u03b1\u03c5\u03c4\u03bf\u03cd \u03c4\u03bf\u03c5 control.<\/p>\n<div class=\"td-decision-band\">\n<p class=\"td-decision-kicker\">\u039a\u03b1\u03bd\u03cc\u03bd\u03b1\u03c2 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae\u03c2 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2<\/p>\n<p class=\"td-decision-title\">\u039c\u03b7\u03bd \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 bit-widths \u03bc\u03b5 \u03ac\u03bd\u03b9\u03c3\u03bf training budget<\/p>\n<p>\u0391\u03bd \u03c4\u03bf 4-bit arm \u03ad\u03c7\u03b5\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf distillation, \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1 \u03ae \u03ac\u03bb\u03bb\u03bf stopping rule, \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c4\u03bf pipeline \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03c4\u03b7\u03bd quantization \u03bc\u03cc\u03bd\u03b7 \u03c4\u03b7\u03c2. \u039a\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 \u03af\u03b4\u03b9\u03bf teacher, tokens, steps, data \u03ba\u03b1\u03b9 evaluation \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03bf\u03b4\u03ce\u03c3\u03b5\u03c4\u03b5 \u03c4\u03b7 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03c3\u03c4\u03b7 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf.<\/p>\n<\/div>\n<p>\u0397 \u03b5\u03c0\u03b9\u03c6\u03cd\u03bb\u03b1\u03be\u03b7 \u03b4\u03b5\u03bd \u03b1\u03ba\u03c5\u03c1\u03ce\u03bd\u03b5\u03b9 \u03c4\u03bf \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1, \u03b1\u03bb\u03bb\u03ac \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c3\u03c9\u03c3\u03c4\u03ac \u03c4\u03b7 \u03b4\u03b9\u03b1\u03c4\u03cd\u03c0\u03c9\u03c3\u03ae \u03c4\u03bf\u03c5. \u0391\u03c5\u03c4\u03cc \u03b5\u03af\u03bd\u03b1\u03b9 \u03bf\u03c5\u03c3\u03b9\u03ce\u03b4\u03b5\u03c2 \u03b3\u03b9\u03b1 E-E-A-T: \u03b7 \u03b1\u03be\u03b9\u03bf\u03c0\u03b9\u03c3\u03c4\u03af\u03b1 \u03b5\u03bd\u03cc\u03c2 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03bf\u03cd \u03ac\u03c1\u03b8\u03c1\u03bf\u03c5 \u03b4\u03b5\u03bd \u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf \u03c0\u03cc\u03c3\u03bf \u03b5\u03bd\u03c4\u03c5\u03c0\u03c9\u03c3\u03b9\u03b1\u03ba\u03cc \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf headline, \u03b1\u03bb\u03bb\u03ac \u03b1\u03c0\u03cc \u03c4\u03bf \u03b1\u03bd \u03be\u03b5\u03c7\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 measurement, inference \u03ba\u03b1\u03b9 \u03c5\u03c0\u03cc\u03b8\u03b5\u03c3\u03b7.<\/p>\n<h2 id=\"mnimi-kostos\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03bf\u03c5\u03bd \u03bc\u03bd\u03ae\u03bc\u03b7, compute \u03ba\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2<\/h2>\n<p>\u03a3\u03cd\u03bc\u03c6\u03c9\u03bd\u03b1 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1, \u03c4\u03bf 4-bit QAH \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03c6\u03bf\u03c1\u03ad\u03c2 \u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 weights \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd BF16 student. \u0395\u03c0\u03b5\u03b9\u03b4\u03ae \u03ad\u03c7\u03b5\u03b9 \u03c4\u03bf \u03bc\u03b9\u03c3\u03cc parameter count \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd 120B teacher, \u03b1\u03bd\u03b1\u03c6\u03ad\u03c1\u03b5\u03c4\u03b1\u03b9 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03bc\u03b9\u03c3\u03cc compute \u03b1\u03bd\u03ac token \u03c3\u03b5 \u03c3\u03c7\u03ad\u03c3\u03b7 \u03bc\u03b5 \u03b5\u03ba\u03b5\u03af\u03bd\u03bf\u03bd. \u039f\u03b9 \u03b1\u03c1\u03b9\u03b8\u03bc\u03bf\u03af \u03b1\u03c6\u03bf\u03c1\u03bf\u03cd\u03bd weights \u03ba\u03b1\u03b9 \u03c4\u03b7 \u03b8\u03b5\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03ae \u03c3\u03c7\u03ad\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf\u03c5 setup\u00b7 \u03b4\u03b5\u03bd \u03b9\u03c3\u03bf\u03b4\u03c5\u03bd\u03b1\u03bc\u03bf\u03cd\u03bd \u03b1\u03c5\u03c4\u03bf\u03bc\u03ac\u03c4\u03c9\u03c2 \u03bc\u03b5 \u03c4\u03b5\u03c4\u03c1\u03b1\u03c0\u03bb\u03ac\u03c3\u03b9\u03b1 \u03c7\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c7\u03c1\u03b7\u03c3\u03c4\u03ce\u03bd \u03ae \u03bc\u03b9\u03c3\u03cc \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc cloud bill.<\/p>\n<p>\u0397 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03bf\u03b9\u03ba\u03bf\u03bd\u03bf\u03bc\u03af\u03b1 \u03b5\u03c0\u03b7\u03c1\u03b5\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc accelerator support \u03b3\u03b9\u03b1 MXFP4, serving framework, batching, concurrency, \u03bc\u03ae\u03ba\u03bf\u03c2 context, KV cache, memory bandwidth, data movement, utilization \u03ba\u03b1\u03b9 observability. \u0388\u03bd\u03b1 format \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03cd\u03c0\u03c9\u03bc\u03b1 \u03c4\u03c9\u03bd weights \u03b1\u03bb\u03bb\u03ac \u03bd\u03b1 \u03bc\u03b7\u03bd \u03c0\u03c1\u03bf\u03c3\u03c6\u03ad\u03c1\u03b5\u03b9 \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7 \u03c3\u03c4\u03bf latency \u03ae \u03c3\u03c4\u03bf throughput \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 hardware stack.<\/p>\n<p>\u0397 \u03b4\u03c5\u03bd\u03b1\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1 \u03bd\u03b1 \u03c7\u03c9\u03c1\u03ad\u03c3\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03b9\u03ba\u03b1\u03bd\u03cc checkpoint \u03c3\u03b5 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf hardware \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03c0\u03b1\u03c1\u2019 \u03cc\u03bb\u03b1 \u03b1\u03c5\u03c4\u03ac \u03bd\u03b1 \u03b1\u03bb\u03bb\u03ac\u03be\u03b5\u03b9 \u03c4\u03bf\u03bd \u03c3\u03c7\u03b5\u03b4\u03b9\u03b1\u03c3\u03bc\u03cc \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae\u03c2: \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b1 replicas, \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03b1\u03c0\u03bf\u03bc\u03cc\u03bd\u03c9\u03c3\u03b7 workloads, on-premises deployment \u03ae \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf \u03b5\u03bc\u03c0\u03cc\u03b4\u03b9\u03bf \u03b3\u03b9\u03b1 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ad\u03c2. \u0397 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03c3\u03c5\u03bd\u03b4\u03ad\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03bf \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc\u03c4\u03b5\u03c1\u03bf \u03b5\u03c1\u03ce\u03c4\u03b7\u03bc\u03b1 \u03c4\u03bf\u03c5 <a href=\"https:\/\/twodots.gr\/poio-ai-montelo-axizei-gia-tin-epicheirisi\/\">\u03c0\u03bf\u03b9\u03bf AI \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03b1\u03be\u03af\u03b6\u03b5\u03b9 \u03b3\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7<\/a> \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03bc\u03b5 \u03ad\u03bd\u03b1 \u03bc\u03cc\u03bd\u03b9\u03bc\u03bf \u03c3\u03c4\u03bf\u03af\u03c7\u03b7\u03bc\u03b1 \u03c3\u03b5 \u03ad\u03bd\u03b1 checkpoint.<\/p>\n<h2 id=\"production-evaluation\">\u03a0\u03ce\u03c2 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03ad\u03bd\u03b1 4-bit \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c3\u03b5 production<\/h2>\n<p>\u039c\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b1 e-commerce, marketing, development \u03ae customer support \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c6\u03b1\u03c3\u03af\u03c3\u03b5\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd \u03bc\u03ad\u03c3\u03bf \u03cc\u03c1\u03bf \u03b5\u03bd\u03bd\u03ad\u03b1 benchmarks. \u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b1\u03bd\u03c4\u03b9\u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b5\u03c5\u03c4\u03b9\u03ba\u03cc evaluation set \u03bc\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03b1\u03c6\u03ad\u03c2 \u03c0\u03c1\u03bf\u03ca\u03cc\u03bd\u03c4\u03c9\u03bd, \u03c4\u03b1\u03be\u03b9\u03bd\u03bf\u03bc\u03ae\u03c3\u03b5\u03b9\u03c2, support \u03b5\u03c1\u03c9\u03c4\u03ae\u03bc\u03b1\u03c4\u03b1, tool calls, policy constraints, \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1 \u03ba\u03b1\u03b9 \u03c4\u03b9\u03c2 \u03b3\u03bb\u03ce\u03c3\u03c3\u03b5\u03c2 \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd \u03bf\u03b9 \u03c0\u03b5\u03bb\u03ac\u03c4\u03b5\u03c2.<\/p>\n<p>\u0397 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03c1\u03ac \u03bc\u03b1\u03b6\u03af \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1: task success, human preference \u03ae defect rate, latency \u03c3\u03c4\u03bf p50 \u03ba\u03b1\u03b9 p95, throughput, peak memory, \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03af\u03b5\u03c2 \u03c3\u03b5 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf context, \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03b1\u03bd\u03ac \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03cc \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc\u03c4\u03b7\u03c4\u03b1 \u03bc\u03b5\u03c4\u03ac \u03c4\u03bf training peak. \u03a3\u03b5 custom kernels \u03ae serving optimizations, \u03b7 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ae \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03ad\u03c3\u03b1 \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u2014 \u03cc\u03c0\u03c9\u03c2 \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03c4\u03bf \u03ac\u03c1\u03b8\u03c1\u03bf \u03b3\u03b9\u03b1 \u03c4\u03bf <a href=\"https:\/\/twodots.gr\/llm4llm-gpu-kernel-dokimi-pragmatiko-montelo\/\">GPU kernel \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 end-to-end validation<\/a>.<\/p>\n<p>\u0391\u03bd \u03ad\u03bd\u03b1 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf checkpoint \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03c4\u03b9\u03c2 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b5\u03c2 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ad\u03c2, \u03c4\u03bf \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03cc. \u0391\u03bd \u03c7\u03ac\u03bd\u03b5\u03b9 \u03c3\u03b5 \u03c3\u03c0\u03ac\u03bd\u03b9\u03b1 \u03b1\u03bb\u03bb\u03ac \u03c5\u03c8\u03b7\u03bb\u03bf\u03cd \u03c1\u03af\u03c3\u03ba\u03bf\u03c5 cases, \u03b7 \u03bc\u03ad\u03c3\u03b7 \u03b2\u03b1\u03b8\u03bc\u03bf\u03bb\u03bf\u03b3\u03af\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03ba\u03c1\u03cd\u03c0\u03c4\u03b5\u03b9 \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2. \u0393\u03b9\u03b1 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03ba\u03b1\u03c4\u03b7\u03b3\u03bf\u03c1\u03af\u03b5\u03c2 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1\u03c2, \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03c1\u03bf\u03c4\u03b9\u03bc\u03cc\u03c4\u03b5\u03c1\u03bf <a href=\"https:\/\/twodots.gr\/llm-routing-latency-accuracy-cost\/\">model routing \u03bc\u03b5 \u03b5\u03c0\u03af\u03b3\u03bd\u03c9\u03c3\u03b7 latency, \u03b1\u03ba\u03c1\u03af\u03b2\u03b5\u03b9\u03b1\u03c2 \u03ba\u03b1\u03b9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c5\u03c2<\/a> \u03b1\u03bd\u03c4\u03af \u03b3\u03b9\u03b1 \u03bc\u03af\u03b1 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7.<\/p>\n<h2 id=\"epta-vimata\">\u03a0\u03b9\u03bb\u03bf\u03c4\u03b9\u03ba\u03ae \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae \u03c3\u03b5 \u03b5\u03c0\u03c4\u03ac \u03b2\u03ae\u03bc\u03b1\u03c4\u03b1<\/h2>\n<p>\u0397 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c3\u03c4\u03b5\u03c1\u03b7 \u03c5\u03b9\u03bf\u03b8\u03ad\u03c4\u03b7\u03c3\u03b7 \u03be\u03b5\u03ba\u03b9\u03bd\u03ac \u03bc\u03b5 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf pilot, \u03cc\u03c7\u03b9 \u03bc\u03b5 \u03ac\u03bc\u03b5\u03c3\u03bf migration \u03cc\u03bb\u03c9\u03bd \u03c4\u03c9\u03bd workloads. \u039a\u03ac\u03b8\u03b5 \u03b2\u03ae\u03bc\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03c4\u03b9\u03ba\u03cc \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03bf \u03c0\u03bf\u03c5 \u03b5\u03c0\u03b9\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03c3\u03c4\u03b7\u03bd \u03bf\u03bc\u03ac\u03b4\u03b1 \u03bd\u03b1 \u03c3\u03c5\u03bd\u03b5\u03c7\u03af\u03c3\u03b5\u03b9, \u03bd\u03b1 \u03b4\u03b9\u03bf\u03c1\u03b8\u03ce\u03c3\u03b5\u03b9 \u03ae \u03bd\u03b1 \u03c3\u03c4\u03b1\u03bc\u03b1\u03c4\u03ae\u03c3\u03b5\u03b9.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">\u0391\u03c0\u03cc \u03c4\u03bf benchmark \u03c3\u03b5 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03bf 4-bit deployment<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 1<\/span><strong>\u039f\u03c1\u03af\u03c3\u03c4\u03b5 \u03c4\u03bf production workload<\/strong>\n<p>\u039a\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 input, output, \u03b5\u03c1\u03b3\u03b1\u03bb\u03b5\u03af\u03b1, \u03b3\u03bb\u03ce\u03c3\u03c3\u03b5\u03c2, context length, concurrency \u03ba\u03b1\u03b9 acceptance criteria \u03b3\u03b9\u03b1 \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc business task.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 2<\/span><strong>\u039a\u03bb\u03b5\u03b9\u03b4\u03ce\u03c3\u03c4\u03b5 matched checkpoints<\/strong>\n<p>\u03a3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 BF16 \u03ba\u03b1\u03b9 MXFP4 \u03bc\u03b5 \u03af\u03b4\u03b9\u03bf teacher, \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1, training tokens, steps \u03ba\u03b1\u03b9 stopping rule \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03bc\u03b7\u03bd \u03bc\u03c0\u03b5\u03c1\u03b4\u03b5\u03cd\u03b5\u03c4\u03b1\u03b9 \u03b7 quantization \u03bc\u03b5 \u03b5\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd compute.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 3<\/span><strong>\u03a7\u03c4\u03af\u03c3\u03c4\u03b5 domain evaluation set<\/strong>\n<p>\u03a7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03c4\u03b5 \u03b1\u03bd\u03c4\u03b9\u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b5\u03c5\u03c4\u03b9\u03ba\u03ac \u03ba\u03b1\u03b9 \u03b4\u03cd\u03c3\u03ba\u03bf\u03bb\u03b1 cases \u03b1\u03c0\u03cc \u03c4\u03bf \u03b4\u03b9\u03ba\u03cc \u03c3\u03b1\u03c2 workflow, \u03bc\u03b1\u03b6\u03af \u03bc\u03b5 edge cases, safety failures \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03bf quality review.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 4<\/span><strong>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 end-to-end \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae<\/strong>\n<p>\u039a\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 p50\/p95 latency, throughput, peak memory, KV cache, utilization \u03ba\u03b1\u03b9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c3\u03c4\u03bf hardware \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf serving framework \u03c0\u03bf\u03c5 \u03b8\u03b1 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b7\u03b8\u03bf\u03cd\u03bd.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 5<\/span><strong>\u0395\u03bb\u03ad\u03b3\u03be\u03c4\u03b5 long context \u03ba\u03b1\u03b9 tool use<\/strong>\n<p>\u0394\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b5 \u03c4\u03b7 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ac \u03c3\u03c4\u03b1 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03bc\u03ae\u03ba\u03b7 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03af\u03b1\u03c2, \u03c3\u03b5 tool calls \u03ba\u03b1\u03b9 \u03c3\u03b5 \u03c0\u03bf\u03bb\u03c5\u03c3\u03c4\u03b1\u03b4\u03b9\u03b1\u03ba\u03ac tasks\u00b7 \u03c4\u03bf nominal context window \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03af \u03b5\u03b3\u03b3\u03cd\u03b7\u03c3\u03b7 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 6<\/span><strong>\u03a4\u03c1\u03ad\u03be\u03c4\u03b5 canary \u03bc\u03b5 rollback<\/strong>\n<p>\u0394\u03c1\u03bf\u03bc\u03bf\u03bb\u03bf\u03b3\u03ae\u03c3\u03c4\u03b5 \u03bc\u03b9\u03ba\u03c1\u03cc \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03cc \u03c7\u03b1\u03bc\u03b7\u03bb\u03bf\u03cd \u03c1\u03af\u03c3\u03ba\u03bf\u03c5 \u03b5\u03c1\u03b3\u03b1\u03c3\u03b9\u03ce\u03bd, \u03ba\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 observability \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03bf gate \u03ba\u03b1\u03b9 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c8\u03c4\u03b5 \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03b7 \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03bf\u03c6\u03ae \u03c3\u03c4\u03bf \u03c0\u03c1\u03bf\u03b7\u03b3\u03bf\u03cd\u03bc\u03b5\u03bd\u03bf checkpoint.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u0421\u0442\u044a\u043f\u043a\u0430 7<\/span><strong>\u0391\u03c0\u03bf\u03c6\u03b1\u03c3\u03af\u03c3\u03c4\u03b5 \u03b1\u03bd\u03ac \u03ba\u03b1\u03c4\u03b7\u03b3\u03bf\u03c1\u03af\u03b1 task<\/strong>\n<p>\u039a\u03bb\u03b9\u03bc\u03b1\u03ba\u03ce\u03c3\u03c4\u03b5 \u03bc\u03cc\u03bd\u03bf \u03cc\u03c0\u03bf\u03c5 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1, latency \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9. \u0393\u03b9\u03b1 \u03b1\u03c0\u03b1\u03b9\u03c4\u03b7\u03c4\u03b9\u03ba\u03ac tasks \u03ba\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ae \u03b5\u03c6\u03b1\u03c1\u03bc\u03cc\u03c3\u03c4\u03b5 routing.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<p>\u03a4\u03bf pilot \u03b4\u03b5\u03bd \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03be\u03b5\u03ba\u03b9\u03bd\u03ae\u03c3\u03b5\u03b9 \u03bc\u03b5 \u03b9\u03b4\u03b9\u03cc\u03ba\u03c4\u03b7\u03c4\u03bf training pipeline. \u039c\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c1\u03c7\u03af\u03c3\u03b5\u03b9 \u03c9\u03c2 structured evaluation \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ce\u03bd checkpoints \u03ba\u03b1\u03b9 hardware \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ce\u03bd. \u0391\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03c7\u03b8\u03b5\u03af \u03cc\u03c4\u03b9 \u03b7 quantization \u03ad\u03c7\u03b5\u03b9 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2, \u03c4\u03cc\u03c4\u03b5 \u03b4\u03b9\u03ba\u03b1\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03b7 \u03b5\u03c0\u03ad\u03bd\u03b4\u03c5\u03c3\u03b7 \u03c3\u03b5 distillation, custom serving \u03ba\u03b1\u03b9 automation.<\/p>\n<h2 id=\"symperasma\">\u0397 \u03bf\u03c5\u03c3\u03af\u03b1 \u03b3\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7<\/h2>\n<p>\u03a4\u03bf Quantization-Aware Healing \u03c0\u03b1\u03c1\u03bf\u03c5\u03c3\u03b9\u03ac\u03b6\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03ba\u03b1\u03b8\u03b1\u03c1\u03ae \u03b9\u03b4\u03ad\u03b1: \u03cc\u03c4\u03b1\u03bd \u03ad\u03bd\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ad\u03c7\u03b5\u03b9 \u03ae\u03b4\u03b7 \u03c5\u03c0\u03bf\u03c3\u03c4\u03b5\u03af \u03b4\u03bf\u03bc\u03b9\u03ba\u03ae \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7, \u03b7 quantized \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03ac\u03b8\u03b5\u03b9 \u03be\u03b1\u03bd\u03ac \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd \u03b1\u03c1\u03c7\u03b9\u03ba\u03cc, \u03b1\u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03c4\u03bf teacher \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1 \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf checkpoint \u03bc\u03b5 \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf ceiling. \u03a4\u03b1 \u03b4\u03b7\u03bc\u03bf\u03c3\u03b9\u03b5\u03c5\u03bc\u03ad\u03bd\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03b4\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03cc\u03c4\u03b9 \u03b1\u03c5\u03c4\u03ae \u03b7 \u03b4\u03b9\u03b1\u03b4\u03b9\u03ba\u03b1\u03c3\u03af\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b4\u03ce\u03c3\u03b5\u03b9 \u03ad\u03bd\u03b1 4-bit deployment artifact \u03c0\u03bf\u03c5 \u03c5\u03c0\u03b5\u03c1\u03b2\u03b1\u03af\u03bd\u03b5\u03b9 \u03c4\u03bf BF16 \u03c3\u03b7\u03bc\u03b5\u03af\u03bf \u03b5\u03b9\u03c3\u03cc\u03b4\u03bf\u03c5 \u03c4\u03bf\u03c5 \u03c3\u03b5 \u03c0\u03bf\u03bb\u03bb\u03ad\u03c2 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2.<\/p>\n<p>\u0397 \u03bc\u03b5\u03b8\u03bf\u03b4\u03bf\u03bb\u03bf\u03b3\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03c6\u03cd\u03bb\u03b1\u03be\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03be\u03af\u03c3\u03bf\u03c5 \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ae \u03bc\u03b5 \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1. \u0394\u03b5\u03bd \u03ad\u03c7\u03b5\u03b9 \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03c7\u03b8\u03b5\u03af \u03cc\u03c4\u03b9 \u03b7 quantization \u03b4\u03b7\u03bc\u03b9\u03bf\u03cd\u03c1\u03b3\u03b7\u03c3\u03b5 \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03b7 \u03c4\u03b7\u03c2 \u03c4\u03b7\u03bd \u03b5\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd \u03b1\u03ba\u03c1\u03af\u03b2\u03b5\u03b9\u03b1. \u0388\u03c7\u03b5\u03b9 \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03c7\u03b8\u03b5\u03af, \u03c3\u03c4\u03bf \u03b1\u03bd\u03b1\u03c6\u03b5\u03c1\u03cc\u03bc\u03b5\u03bd\u03bf case study, \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1 \u03c3\u03c5\u03bc\u03c0\u03b9\u03b5\u03c3\u03bc\u03ad\u03bd\u03bf 4-bit checkpoint \u03bc\u03b5 \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c4\u03b7 distillation \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf BF16 checkpoint \u03c0\u03bf\u03c5 \u03bc\u03c0\u03ae\u03ba\u03b5 \u03c3\u03c4\u03bf \u03c3\u03c4\u03ac\u03b4\u03b9\u03bf \u03b1\u03c5\u03c4\u03cc.<\/p>\n<p>\u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7, \u03c4\u03bf QAH \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03ac\u03b4\u03b5\u03b9\u03b1 \u03b3\u03b9\u03b1 \u03b3\u03b5\u03bd\u03af\u03ba\u03b5\u03c5\u03c3\u03b7. \u0395\u03af\u03bd\u03b1\u03b9 \u03c0\u03c1\u03cc\u03c3\u03ba\u03bb\u03b7\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03b1\u03c5\u03c3\u03c4\u03b7\u03c1\u03cc \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1: matched controls, domain evals, end-to-end serving \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03b1\u03bd\u03ac workload. \u0397 \u03b1\u03be\u03af\u03b1 \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03c0\u03bf\u03c5 \u03c0\u03b5\u03c1\u03bd\u03ac \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03b1 \u03c4\u03b1 \u03b4\u03b9\u03ba\u03ac \u03c3\u03b1\u03c2 quality gates \u2014 \u03cc\u03c7\u03b9 \u03c3\u03c4\u03bf \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf bit-width \u03c9\u03c2 \u03c3\u03cd\u03bd\u03b8\u03b7\u03bc\u03b1.<\/p>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<p class=\"td-service-cta-kicker\">AI \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae \u03bc\u03b5 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03b7 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7<\/p>\n<p class=\"td-service-cta-title\">\u0394\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b5 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03bc\u03b9\u03ba\u03c1\u03cd\u03bd\u03b5\u03c4\u03b5 \u03c4\u03b1 quality gates<\/p>\n<p>\u0397 TWO DOTS \u03c7\u03b1\u03c1\u03c4\u03bf\u03b3\u03c1\u03b1\u03c6\u03b5\u03af workload, evaluation set, integrations, human review \u03ba\u03b1\u03b9 fallback \u03ce\u03c3\u03c4\u03b5 \u03ad\u03bd\u03b1 4-bit AI deployment \u03bd\u03b1 \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1, latency \u03ba\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03cc \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2 \u03c0\u03c1\u03b9\u03bd \u03ba\u03bb\u03b9\u03bc\u03b1\u03ba\u03c9\u03b8\u03b5\u03af.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">\u0394\u03b5\u03af\u03c4\u03b5 \u03bb\u03cd\u03c3\u03b5\u03b9\u03c2 \u0391\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03ce\u03bd \u0395\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03ae\u03c3\u03b5\u03c9\u03bd &amp; AI<\/a><\/div>\n<\/div>\n<\/section>\n<section id=\"sychnes-erotiseis\" class=\"td-faq-section\">\n<div class=\"td-faq\">\n<p class=\"td-faq-heading\">\u0427\u0435\u0441\u0442\u043e \u0437\u0430\u0434\u0430\u0432\u0430\u043d\u0438 \u0432\u044a\u043f\u0440\u043e\u0441\u0438<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf Quantization-Aware Healing;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03af\u03bd\u03b1\u03b9 \u03bc\u03b9\u03b1 \u03c3\u03c5\u03bd\u03c4\u03b1\u03b3\u03ae distillation \u03b3\u03b9\u03b1 structurally compressed \u03ba\u03b1\u03b9 quantized LLMs. \u039f 4-bit student \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03bc\u03b9\u03bc\u03b5\u03af\u03c4\u03b1\u03b9 \u03c4\u03b1 output logits \u03c4\u03bf\u03c5 \u03b1\u03c1\u03c7\u03b9\u03ba\u03bf\u03cd, \u03b1\u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03c4\u03bf\u03c5 teacher \u03b1\u03bd\u03c4\u03af \u03bd\u03b1 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf BF16 checkpoint.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03bf 4-bit \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf \u03b1\u03c1\u03c7\u03b9\u03ba\u03cc GPT-OSS 120B;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ac. \u0399\u03c3\u03bf\u03c6\u03ac\u03c1\u03b9\u03c3\u03b5 \u03ae \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf \u03c3\u03c5\u03bc\u03c0\u03b9\u03b5\u03c3\u03bc\u03ad\u03bd\u03bf 60B BF16 checkpoint \u03c3\u03b5 7 \u03b1\u03c0\u03cc 9 benchmarks. \u0388\u03c6\u03c4\u03b1\u03c3\u03b5 \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ac \u03c4\u03bf\u03bd 120B teacher \u03c3\u03c4\u03bf LiveCodeBench, \u03b1\u03bb\u03bb\u03ac \u03c0\u03b1\u03c1\u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03c0\u03af\u03c3\u03c9 \u03c3\u03c4\u03b9\u03c2 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03ac\u03bb\u03bb\u03b5\u03c2 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf headline \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b5\u03c0\u03b9\u03c6\u03cd\u03bb\u03b1\u03be\u03b7;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf 60B MXFP4 \u03ad\u03bb\u03b1\u03b2\u03b5 \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c4\u03bf distillation pass \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd 120B teacher, \u03b5\u03bd\u03ce \u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03c1\u03b9\u03bd\u03cc\u03bc\u03b5\u03bd\u03bf 60B BF16 \u03cc\u03c7\u03b9. \u03a7\u03c9\u03c1\u03af\u03c2 matched BF16 control \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03bc\u03bf\u03bd\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b7 \u03b5\u03c0\u03af\u03b4\u03c1\u03b1\u03c3\u03b7 \u03c4\u03b7\u03c2 quantization \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b5\u03c0\u03af\u03b4\u03c1\u03b1\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf\u03c5 training.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac QAH \u03ba\u03b1\u03b9 QAT;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf QAH \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af KL distillation \u03c0\u03c1\u03bf\u03c2 frozen teacher distribution. \u03a4\u03bf QAT \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af task loss \u03bc\u03b5 fake quantization. \u03a3\u03c4\u03bf \u03b4\u03b7\u03bc\u03bf\u03c3\u03b9\u03b5\u03c5\u03bc\u03ad\u03bd\u03bf 9B \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03b5\u03af\u03c7\u03b1\u03bd \u03c0\u03b1\u03c1\u03cc\u03bc\u03bf\u03b9\u03bf peak, \u03b1\u03bb\u03bb\u03ac \u03c4\u03bf QAH \u03c4\u03bf \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 \u03bd\u03c9\u03c1\u03af\u03c4\u03b5\u03c1\u03b1 \u03ba\u03b1\u03b9 \u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03c0\u03b9\u03bf \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 MXFP4;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03af\u03bd\u03b1\u03b9 4-bit microscaling floating-point format. \u03a3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf pipeline \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af\u03c4\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03c4\u03b1 weights \u03c4\u03bf\u03c5 60B student, \u03b5\u03bd\u03ce \u03c4\u03bf \u03b5\u03c0\u03af\u03c3\u03b7\u03bc\u03bf GPT-OSS \u03ad\u03c7\u03b5\u03b9 \u03ae\u03b4\u03b7 \u03c4\u03b7 \u03bc\u03b5\u03b3\u03ac\u03bb\u03b7 \u03c0\u03bb\u03b5\u03b9\u03bf\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03c9\u03bd MoE weights \u03c3\u03b5 MXFP4.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 4-bit \u03cc\u03c4\u03b9 \u03c4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c0\u03ad\u03c6\u03c4\u03b5\u03b9 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03c6\u03bf\u03c1\u03ad\u03c2;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9. \u0397 \u03b1\u03bd\u03b1\u03c6\u03bf\u03c1\u03ac \u03b3\u03b9\u03b1 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03c6\u03bf\u03c1\u03ad\u03c2 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c4\u03b5\u03c1\u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03b1\u03c6\u03bf\u03c1\u03ac \u03c4\u03b1 weights \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 BF16. \u03a4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc hardware support, serving, batching, context, KV cache, utilization \u03ba\u03b1\u03b9 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2 \u03c3\u03b5 long-context \u03c7\u03c1\u03ae\u03c3\u03b7;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf QAH \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c4\u03b7\u03ba\u03b5 \u03bc\u03b5 sequence length 32k \u03bc\u03ad\u03c3\u03c9 offline top-k logits \u03ba\u03b1\u03b9 chunked KL loss, \u03b5\u03bd\u03ce \u03c4\u03bf \u03bc\u03b5\u03b3\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03ba\u03ad\u03c1\u03b4\u03bf\u03c2 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 BF16 \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03c3\u03c4\u03b7\u03ba\u03b5 \u03c3\u03c4\u03bf AA-LCR. \u0397 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03bb\u03b5\u03b3\u03c7\u03b8\u03b5\u03af \u03c3\u03c4\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b1 \u03bc\u03ae\u03ba\u03b7 context \u03ba\u03ac\u03b8\u03b5 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc \u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03bf \u03b2\u03ae\u03bc\u03b1 \u03b3\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b5\u03af\u03b1;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039d\u03b1 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03b9 matched checkpoints \u03c3\u03b5 \u03b1\u03bd\u03c4\u03b9\u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b5\u03c5\u03c4\u03b9\u03ba\u03cc evaluation set \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc serving stack, \u03bc\u03b5\u03c4\u03c1\u03ce\u03bd\u03c4\u03b1\u03c2 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1, latency, throughput, \u03bc\u03bd\u03ae\u03bc\u03b7, \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03ba\u03b1\u03b9 high-risk edge cases \u03c0\u03c1\u03b9\u03bd \u03b5\u03c0\u03b9\u03bb\u03ad\u03be\u03b5\u03b9 production model.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">\u0418\u0437\u0442\u043e\u0447\u043d\u0438\u0446\u0438<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.20953\" target=\"_blank\" rel=\"noopener\">Multiverse Computing \u2014 Quantization-Aware Healing: A Practical Recipe for Recovering Compressed, 4-Bit LLMs<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/blog\/MultiverseComputingCAI\/quantization-aware-healing\" target=\"_blank\" rel=\"noopener\">Multiverse Computing \u03c3\u03c4\u03bf Hugging Face \u2014 QAH, \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03b4\u03b7\u03bc\u03cc\u03c3\u03b9\u03b1 \u03c3\u03c5\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03c4\u03bf missing control<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.03796\" target=\"_blank\" rel=\"noopener\">Multiverse Computing \u2014 Efficient Knowledge Distillation for LLMs: Offline Top-K Logits and a Fused Chunked KL Loss<\/a><\/li>\n<li><a href=\"https:\/\/openai.com\/index\/gpt-oss-model-card\/\" target=\"_blank\" rel=\"noopener\">OpenAI \u2014 gpt-oss-120b &amp; gpt-oss-20b Model Card<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>\u03a4\u03bf Quantization-Aware Healing \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b5 4-bit GPT-OSS \u03c3\u03b5 7 \u03b1\u03c0\u03cc 9 benchmarks. \u03a4\u03b9 \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1 \u03bc\u03bd\u03ae\u03bc\u03b7, \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2, controls \u03ba\u03b1\u03b9 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 AI deployment.<\/p>","protected":false},"author":1,"featured_media":105411,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","content-type":"","footnotes":""},"categories":[199],"tags":[8988,20408,7477,20790,20789],"class_list":["post-98123","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-ai-infrastructure","tag-ai-models","tag-machine-learning","tag-quantization","tag-quantization-aware-healing"],"blocksy_meta":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":8}},"_links":{"self":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/98123","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/comments?post=98123"}],"version-history":[{"count":1,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/98123\/revisions"}],"predecessor-version":[{"id":105412,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/98123\/revisions\/105412"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/media\/105411"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/media?parent=98123"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/categories?post=98123"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/tags?post=98123"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}