{"id":2649,"date":"2024-08-23T04:16:56","date_gmt":"2024-08-23T04:16:56","guid":{"rendered":"https:\/\/tastycounter.net\/index.php\/2024\/08\/23\/microsoft-phi-3-5-ra-mat-mo-hinh-ai-giau-tinh-canh-tranh-hon\/"},"modified":"2024-08-23T04:16:56","modified_gmt":"2024-08-23T04:16:56","slug":"microsoft-phi-3-5-ra-mat-mo-hinh-ai-giau-tinh-canh-tranh-hon","status":"publish","type":"post","link":"https:\/\/tastycounter.net\/index.php\/2024\/08\/23\/microsoft-phi-3-5-ra-mat-mo-hinh-ai-giau-tinh-canh-tranh-hon\/","title":{"rendered":"Microsoft Phi-3.5 ra m\u1eaft: M\u00f4 h\u00ecnh AI gi\u00e0u t\u00ednh c\u1ea1nh tranh h\u01a1n"},"content":{"rendered":"<\/p>\n<div class=\"content-detail textview\">\n<div class=\"audio\"><audio controls><\/audio><\/div>\n<p>Microsoft v\u1eeba ch\u00ednh th\u1ee9c c\u00f4ng b\u1ed1 ph\u00e1t h\u00e0nh lo\u1ea1t m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef nh\u1ecf (Small Language Model) m\u1edbi c\u00f3 t\u00ean Phi-3.5 m\u1edbi, bao g\u1ed3m ba bi\u1ebfn th\u1ec3 Phi-3.5-vision, Phi-3.5-MoE v\u00e0 Phi-3.5-mini. C\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef AI g\u1ecdn nh\u1eb9 n\u00e0y \u0111\u01b0\u1ee3c x\u00e2y d\u1ef1ng d\u1ef1a tr\u00ean d\u1eef li\u1ec7u t\u1ed5ng h\u1ee3p v\u00e0 c\u00e1c trang web c\u00f4ng khai \u0111\u00e3 \u0111\u01b0\u1ee3c l\u1ecdc, h\u1ed7 tr\u1ee3 \u0111\u1ed9 d\u00e0i ng\u1eef c\u1ea3nh 128K token. T\u1ea5t c\u1ea3 c\u00e1c m\u00f4 h\u00ecnh Phi-3.5 m\u1edbi hi\u1ec7n \u0111\u00e3 kh\u1ea3 d\u1ee5ng tr\u00ean Hugging Face theo gi\u1ea5y ph\u00e9p MIT.<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" style=\"margin-left:auto;margin-right:auto\" src=\"https:\/\/st.quantrimang.com\/photos\/image\/holder.png\" alt=\"Microsoft Phi-3.5\" width=\"400\" height=\"224\" class=\"lazy\" data-src=\"https:\/\/st.quantrimang.com\/photos\/image\/2024\/08\/22\/microsoft-phi-3-51.jpg\"><\/p>\n<h2>Phi-3.5-MoE: S\u1ef1 k\u1ebft h\u1ee3p mang nhi\u1ec1u t\u00ednh \u0111\u1ed9t ph\u00e1<\/h2>\n<p>Phi-3.5-MoE n\u1ed5i b\u1eadt l\u00e0 m\u00f4 h\u00ecnh \u0111\u1ea7u ti\u00ean trong h\u1ecd Phi c\u1ee7a Microsoft c\u00f3 th\u1ec3 t\u1eadn d\u1ee5ng c\u00f4ng ngh\u1ec7 Mixture of Experts (MoE). M\u00f4 h\u00ecnh MoE 16 x 3,8 t\u1ef7 tham s\u1ed1 n\u00e0y ch\u1ec9 k\u00edch ho\u1ea1t 6,6 t\u1ef7 tham s\u1ed1 v\u00e0 \u0111\u01b0\u1ee3c \u0111\u00e0o t\u1ea1o tr\u00ean 4,9T token s\u1eed d\u1ee5ng 512 h\u1ec7 th\u1ed1ng GPU H100. Trong c\u00e1c chu\u1ea9n AI ph\u1ed5 bi\u1ebfn hi\u1ec7n nay, Phi-3.5-MoE v\u01b0\u1ee3t tr\u1ed9i h\u01a1n Llama-3.1 8B, Gemma-2-9B v\u00e0 Gemini-1.5-Flash, v\u00e0 ti\u1ec7m c\u1eadn v\u1edbi m\u00f4 h\u00ecnh d\u1eabn \u0111\u1ea7u hi\u1ec7n t\u1ea1i l\u00e0 GPT-4o-mini.<\/p>\n<h2>Phi-3.5-mini: G\u1ecdn nh\u1eb9 v\u00e0 M\u1ea1nh m\u1ebd<\/h2>\n<p>Phi-3.5-mini l\u00e0 m\u00f4 h\u00ecnh 3.8 t\u1ef7 tham s\u1ed1, v\u01b0\u1ee3t tr\u1ed9i h\u01a1n Llama3.1 8B hay Mistral 7B, v\u00e0 th\u1eadm ch\u00ed c\u00f2n c\u1ea1nh tranh v\u1edbi Mistral NeMo 12B. N\u00f3 \u0111\u01b0\u1ee3c \u0111\u00e0o t\u1ea1o tr\u00ean 3.4T token s\u1eed d\u1ee5ng 512 GPU H100. V\u1edbi ch\u1ec9 3.8B tham s\u1ed1 ho\u1ea1t \u0111\u1ed9ng, m\u00f4 h\u00ecnh n\u00e0y c\u00f3 kh\u1ea3 n\u0103ng c\u1ea1nh tranh tr\u00ean c\u00e1c t\u00e1c v\u1ee5 \u0111a ng\u00f4n ng\u1eef so v\u1edbi c\u00e1c LLM c\u00f3 nhi\u1ec1u tham s\u1ed1 ho\u1ea1t \u0111\u1ed9ng h\u01a1n. Ngo\u00e0i ra, Phi-3.5-mini hi\u1ec7n h\u1ed7 tr\u1ee3 \u0111\u1ed9 d\u00e0i ng\u1eef c\u1ea3nh 128K, trong khi \u0111\u1ed1i th\u1ee7 ch\u00ednh Gemma-2 ch\u1ec9 h\u1ed7 tr\u1ee3 8K.<\/p>\n<h2>Phi-3.5-vision: N\u00e2ng cao kh\u1ea3 n\u0103ng x\u1eed l\u00fd h\u00ecnh \u1ea3nh \u0111a khung<\/h2>\n<p>Phi-3.5-vision l\u00e0 m\u00f4 h\u00ecnh 4.2 t\u1ef7 tham s\u1ed1 \u0111\u01b0\u1ee3c \u0111\u00e0o t\u1ea1o tr\u00ean 500B token s\u1eed d\u1ee5ng 256 GPU A100. M\u00f4 h\u00ecnh n\u00e0y hi\u1ec7n h\u1ed7 tr\u1ee3 kh\u1ea3 n\u0103ng hi\u1ec3u v\u00e0 suy lu\u1eadn h\u00ecnh \u1ea3nh \u0111a khung (multi-frame). Phi-3.5-vision \u0111\u00e3 c\u1ea3i thi\u1ec7n hi\u1ec7u su\u1ea5t tr\u00ean MMMU (t\u1eeb 40.2 l\u00ean 43.0), MMBench (t\u1eeb 80.5 l\u00ean 81.9) v\u00e0 chu\u1ea9n x\u1eed l\u00fd t\u00e0i li\u1ec7u TextVQA (t\u1eeb 70.9 l\u00ean 72.0).<\/p>\n<p>Microsoft c\u00f3 k\u1ebf ho\u1ea1ch chia s\u1ebb th\u00eam th\u00f4ng tin chi ti\u1ebft v\u1ec1 d\u00f2ng m\u00f4 h\u00ecnh Phi-3.5 trong th\u00e1ng n\u00e0y, ch\u1ee7 y\u1ebfu gi\u1edbi thi\u1ec7u nh\u1eefng ti\u1ebfn b\u1ed9 v\u1ec1 hi\u1ec7u qu\u1ea3 v\u00e0 kh\u1ea3 n\u0103ng c\u1ee7a m\u00f4 h\u00ecnh AI. V\u1edbi tr\u1ecdng t\u00e2m l\u00e0 thi\u1ebft k\u1ebf nh\u1eb9 v\u00e0 hi\u1ec3u bi\u1ebft \u0111a ph\u01b0\u01a1ng th\u1ee9c, d\u00f2ng m\u00f4 h\u00ecnh Phi-3.5 c\u00f3 th\u1ec3 \u0111\u01b0\u1ee3c \u00e1p d\u1ee5ng r\u1ed9ng r\u00e3i h\u01a1n tr\u00ean nhi\u1ec1u \u1ee9ng d\u1ee5ng AI kh\u00e1c nhau.<\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>Microsoft v\u1eeba ch\u00ednh th\u1ee9c c\u00f4ng b\u1ed1 ph\u00e1t h\u00e0nh lo\u1ea1t m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef nh\u1ecf (Small Language Model) m\u1edbi c\u00f3 t\u00ean Phi-3.5 m\u1edbi, bao g\u1ed3m ba bi\u1ebfn th\u1ec3 Phi-3.5-vision, Phi-3.5-MoE v\u00e0 Phi-3.5-mini. C\u00e1c m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef AI g\u1ecdn nh\u1eb9 n\u00e0y \u0111\u01b0\u1ee3c x\u00e2y d\u1ef1ng d\u1ef1a tr\u00ean d\u1eef li\u1ec7u t\u1ed5ng h\u1ee3p v\u00e0 c\u00e1c trang web c\u00f4ng khai [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2649","post","type-post","status-publish","format-standard","hentry","category-khong-phan-loai"],"_links":{"self":[{"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/posts\/2649","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/comments?post=2649"}],"version-history":[{"count":0,"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/posts\/2649\/revisions"}],"wp:attachment":[{"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/media?parent=2649"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/categories?post=2649"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tastycounter.net\/index.php\/wp-json\/wp\/v2\/tags?post=2649"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}