{"id":75810,"date":"2026-04-24T09:06:42","date_gmt":"2026-04-24T02:06:42","guid":{"rendered":"https:\/\/hbbgroup.net\/tencents-new-hy3-ai-model-is-the-most-efficient-chinese-llm-no-ones-talking-about\/"},"modified":"2026-04-24T10:05:01","modified_gmt":"2026-04-24T03:05:01","slug":"tencents-new-hy3-ai-model-is-the-most-efficient-chinese-llm-no-ones-talking-about","status":"publish","type":"post","link":"https:\/\/hbbgroup.net\/vi\/tencents-new-hy3-ai-model-is-the-most-efficient-chinese-llm-no-ones-talking-about\/","title":{"rendered":"Tencent ra m\u1eaft m\u00f4 h\u00ecnh AI Hy3 m\u1edbi: LLM hi\u1ec7u qu\u1ea3 nh\u1ea5t Trung Qu\u1ed1c m\u00e0 \u00edt ng\u01b0\u1eddi nh\u1eafc \u0111\u1ebfn"},"content":{"rendered":"
\n

T\u00f3m t\u1eaft nhanh (In brief)<\/strong><\/h3>\n
    \n
  • Tencent Hy3 preview l\u00e0 m\u00f4 h\u00ecnh Mixture-of-Experts v\u1edbi 295 t\u1ef7 tham s\u1ed1, nh\u01b0ng ch\u1ec9 21 t\u1ef7 tham s\u1ed1 ho\u1ea1t \u0111\u1ed9ng t\u1ea1i m\u1ed9t th\u1eddi \u0111i\u1ec3m, gi\u00fap chi ph\u00ed v\u1eadn h\u00e0nh th\u1ea5p h\u01a1n nhi\u1ec1u \u0111\u1ed1i th\u1ee7 c\u00f9ng ph\u00e2n kh\u00fac.<\/li>\n
  • Tr\u00ean SWE-bench Verified \u2014 benchmark s\u1eeda l\u1ed7i code th\u1ef1c t\u1ebf tr\u00ean GitHub \u2014 model t\u0103ng t\u1eeb 53% (Hy2) l\u00ean 74,4%, t\u01b0\u01a1ng \u0111\u01b0\u01a1ng m\u1ee9c c\u1ea3i thi\u1ec7n 40% so v\u1edbi th\u1ebf h\u1ec7 tr\u01b0\u1edbc.<\/li>\n
  • Model \u0111\u00e3 \u0111\u01b0\u1ee3c tri\u1ec3n khai trong h\u1ec7 sinh th\u00e1i Tencent nh\u01b0 Yuanbao, QQ v\u00e0 Tencent Docs, v\u1edbi API tr\u00ean Tencent Cloud c\u00f3 gi\u00e1 kh\u1edfi \u0111i\u1ec3m kho\u1ea3ng $0,18\/tri\u1ec7u input tokens.<\/li>\n<\/ul>\n
    \n

    N\u1ed9i dung ch\u00ednh<\/strong><\/h3>\n

    Tencent \u0111\u00e3 \u00e2m th\u1ea7m ra m\u1eaft m\u00f4 h\u00ecnh AI m\u1ea1nh nh\u1ea5t c\u1ee7a m\u00ecnh v\u00e0o th\u1ee9 N\u0103m, v\u1edbi c\u00e1c ch\u1ec9 s\u1ed1 benchmark r\u1ea5t \u0111\u00e1ng ch\u00fa \u00fd. Hy3 preview \u2014 model \u0111\u1ea7u ti\u00ean sau khi h\u00e3ng t\u00e1i x\u00e2y d\u1ef1ng to\u00e0n b\u1ed9 h\u1ea1 t\u1ea7ng \u2014 hi\u1ec7n \u0111\u00e3 \u0111\u01b0\u1ee3c open-source tr\u00ean GitHub, Hugging Face v\u00e0 ModelScope.<\/p>\n

    Model c\u0169ng c\u00f3 s\u1eb5n tr\u00ean Tencent Cloud d\u01b0\u1edbi d\u1ea1ng d\u1ecbch v\u1ee5 tr\u1ea3 ph\u00ed.<\/p>\n

    Hy3 s\u1edf h\u1eefu t\u1ed5ng c\u1ed9ng 295 t\u1ef7 tham s\u1ed1 (th\u1ec3 hi\u1ec7n \u0111\u1ed9 r\u1ed9ng tri th\u1ee9c), nh\u01b0ng ch\u1ec9 k\u00edch ho\u1ea1t 21 t\u1ef7 tham s\u1ed1 t\u1ea1i m\u1ed7i l\u1ea7n x\u1eed l\u00fd. \u0110\u00e2y l\u00e0 \u01b0u \u0111i\u1ec3m c\u1ee7a ki\u1ebfn tr\u00fac Mixture-of-Experts \u2014 model s\u1ebd \u0111\u1ecbnh tuy\u1ebfn truy v\u1ea5n \u0111\u1ebfn m\u1ed9t nh\u00f3m \u201cexpert\u201d chuy\u00ean bi\u1ec7t thay v\u00ec ch\u1ea1y to\u00e0n b\u1ed9 m\u1ea1ng c\u00f9ng l\u00fac, gi\u00fap gi\u1ea3m compute, t\u1ed1i \u01b0u chi ph\u00ed nh\u01b0ng v\u1eabn gi\u1eef ch\u1ea5t l\u01b0\u1ee3ng \u0111\u1ea7u ra t\u01b0\u01a1ng \u0111\u01b0\u01a1ng. Model c\u0169ng h\u1ed7 tr\u1ee3 context l\u00ean \u0111\u1ebfn 256.000 tokens, \u0111\u1ee7 \u0111\u1ec3 x\u1eed l\u00fd c\u1ea3 m\u1ed9t cu\u1ed1n ti\u1ec3u thuy\u1ebft trong m\u1ed9t prompt.<\/p>\n

    Tencent cho bi\u1ebft Hy3 \u0111\u01b0\u1ee3c thi\u1ebft k\u1ebf \u0111\u1ec3 c\u00e2n b\u1eb1ng ba y\u1ebfu t\u1ed1 m\u00e0 tr\u01b0\u1edbc \u0111\u00e2y kh\u00f3 \u0111\u1ea1t c\u00f9ng l\u00fac: \u0111\u1ed9 r\u1ed9ng n\u0103ng l\u1ef1c, \u0111\u00e1nh gi\u00e1 trung th\u1ef1c v\u00e0 hi\u1ec7u qu\u1ea3 chi ph\u00ed. Model ti\u1ec1n nhi\u1ec7m Hy2 c\u00f3 h\u01a1n 400 t\u1ef7 tham s\u1ed1, nh\u01b0ng Tencent \u0111\u00e3 ch\u1ee7 \u0111\u1ed9ng gi\u1ea3m xu\u1ed1ng 295 t\u1ef7, cho r\u1eb1ng \u0111\u00e2y l\u00e0 \u201c\u0111i\u1ec3m t\u1ed1i \u01b0u\u201d khi kh\u1ea3 n\u0103ng suy lu\u1eadn \u0111\u1ea1t \u0111\u1ed9 tr\u01b0\u1edfng th\u00e0nh trong khi chi ph\u00ed t\u0103ng th\u00eam kh\u00f4ng c\u00f2n hi\u1ec7u qu\u1ea3.<\/p>\n

    \u0110i\u1ec1u n\u00e0y kh\u00f4ng \u0111\u1ed3ng ngh\u0129a model y\u1ebfu h\u01a1n \u2014 th\u1ef1c t\u1ebf, c\u00e1c m\u00f4 h\u00ecnh \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n t\u1ed1t v\u1edbi \u00edt tham s\u1ed1 h\u01a1n th\u01b0\u1eddng c\u00f3 th\u1ec3 v\u01b0\u1ee3t qua c\u00e1c m\u00f4 h\u00ecnh l\u1edbn nh\u01b0ng k\u00e9m t\u1ed1i \u01b0u.<\/p>\n

    \u1ede m\u1ea3ng coding, b\u01b0\u1edbc ti\u1ebfn l\u00e0 r\u1ea5t r\u00f5 r\u1ec7t. Tr\u00ean SWE-bench Verified \u2014 benchmark ki\u1ec3m tra kh\u1ea3 n\u0103ng s\u1eeda l\u1ed7i th\u1ef1c t\u1ebf tr\u00ean GitHub \u2014 Hy2 \u0111\u1ea1t 53,0% trong khi Hy3 preview \u0111\u1ea1t 74,4%, t\u0103ng kho\u1ea3ng 40% ch\u1ec9 sau m\u1ed9t th\u1ebf h\u1ec7. Th\u00e0nh t\u00edch n\u00e0y \u0111\u01b0a Hy3 ti\u1ec7m c\u1eadn Anthropic Claude Opus 4.6 (80,8%) v\u00e0 v\u01b0\u1ee3t qua c\u00e1c model nh\u01b0 GLM-5 (77,8%) v\u00e0 Kimi-K2.5 (76,8%).<\/p>\n

    Ngo\u00e0i ra, tr\u00ean Terminal-Bench 2.0 \u2014 benchmark \u0111o kh\u1ea3 n\u0103ng th\u1ef1c thi t\u00e1c v\u1ee5 t\u1ef1 \u0111\u1ed9ng trong m\u00f4i tr\u01b0\u1eddng d\u00f2ng l\u1ec7nh \u2014 model c\u0169ng t\u0103ng m\u1ea1nh t\u1eeb 23,2% l\u00ean 54,4%.<\/p>\n<\/div>\n

    \n
    \"\"<\/figure>\n<\/div>\n

    Tuy nhi\u00ean, model n\u00e0y \u0111\u1eb7c bi\u1ec7t \u0111\u00e1ng ch\u00fa \u00fd \u0111\u1ed1i v\u1edbi nh\u1eefng ng\u01b0\u1eddi x\u00e2y d\u1ef1ng h\u1ec7 th\u1ed1ng agent. Agent th\u01b0\u1eddng ph\u1ea3i x\u1eed l\u00fd t\u1eadp h\u1ee3p ch\u1ec9 d\u1eabn ph\u1ee9c t\u1ea1p bao g\u1ed3m memory, skill v\u00e0 c\u00e1c tool call. Ch\u1ec9 c\u1ea7n b\u1ecf s\u00f3t m\u1ed9t y\u1ebfu t\u1ed1 nh\u1ecf c\u0169ng c\u00f3 th\u1ec3 l\u00e0m h\u1ecfng to\u00e0n b\u1ed9 workflow ho\u1eb7c t\u1ea1o ra k\u1ebft qu\u1ea3 k\u00e9m ch\u1ea5t l\u01b0\u1ee3ng. Ch\u00ednh v\u00ec v\u1eady, n\u0103ng l\u1ef1c agentic ng\u00e0y c\u00e0ng tr\u1edf th\u00e0nh y\u1ebfu t\u1ed1 c\u1ed1t l\u00f5i \u0111\u1ed1i v\u1edbi developer AI khi l\u0129nh v\u1ef1c n\u00e0y \u0111ang l\u00e0 xu h\u01b0\u1edbng \u201chot\u201d nh\u1ea5t trong ng\u00e0nh. \u0110\u00e2y c\u0169ng l\u00e0 l\u00fd do model \u0111\u01b0\u1ee3c \u0111\u01b0a l\u00ean Openclaw ngay t\u1eeb \u0111\u1ea7u.<\/p>\n

    \u1ede m\u1ea3ng search v\u00e0 browsing agent \u2014 n\u01a1i model ph\u1ea3i t\u1ef1 thu th\u1eadp, l\u1ecdc v\u00e0 t\u1ed5ng h\u1ee3p th\u00f4ng tin t\u1eeb web m\u00e0 kh\u00f4ng c\u00f3 s\u1ef1 h\u01b0\u1edbng d\u1eabn c\u1ee7a con ng\u01b0\u1eddi \u2014 hi\u1ec7u su\u1ea5t c\u0169ng c\u1ea3i thi\u1ec7n m\u1ea1nh. Tr\u00ean BrowseComp, Hy3 preview \u0111\u1ea1t 67,1% (so v\u1edbi 28,7% c\u1ee7a Hy2). Tr\u00ean WideSearch, model \u0111\u1ea1t 70,2%, v\u01b0\u1ee3t GLM-5 v\u00e0 Kimi-K2.5 nh\u01b0ng v\u1eabn th\u1ea5p h\u01a1n Anthropic Claude Opus 4.6 (77,2%).<\/p>\n

    V\u1ec1 kh\u1ea3 n\u0103ng reasoning, Tencent Hy3 d\u1eabn \u0111\u1ea7u c\u00e1c model Trung Qu\u1ed1c trong k\u1ef3 thi tuy\u1ec3n sinh ti\u1ebfn s\u0129 To\u00e1n c\u1ee7a Tsinghua University (Spring 2026), \u0111\u1ea1t \u0111i\u1ec3m trung b\u00ecnh 88,4 (avg@3). \u0110\u00e2y l\u00e0 b\u00e0i thi th\u1ef1c t\u1ebf, kh\u00f4ng ph\u1ea3i dataset \u0111\u01b0\u1ee3c t\u1ed1i \u01b0u s\u1eb5n \u2014 \u0111\u00fang theo \u0111\u1ecbnh h\u01b0\u1edbng c\u1ee7a Tencent nh\u1eb1m tr\u00e1nh \u201cbenchmark gaming\u201d. Model c\u0169ng \u0111\u1ea1t 87,8 \u0111i\u1ec3m t\u1ea1i CHSBO 2025 (Olympic Sinh h\u1ecdc THPT Trung Qu\u1ed1c), cao nh\u1ea5t trong c\u00e1c model n\u1ed9i \u0111\u1ecba \u1edf h\u1ea1ng m\u1ee5c n\u00e0y.<\/p>\n

    Hy3 preview b\u1eaft \u0111\u1ea7u \u0111\u01b0\u1ee3c train t\u1eeb cu\u1ed1i th\u00e1ng 1\/2026 v\u00e0 ra m\u1eaft v\u00e0o th\u1ee9 N\u0103m \u2014 ch\u01b0a \u0111\u1ebfn 3 th\u00e1ng t\u1eeb khi kh\u1edfi \u0111\u1ea7u \u0111\u1ebfn khi open-source, m\u1ed9t t\u1ed1c \u0111\u1ed9 hi\u1ebfm th\u1ea5y v\u1edbi model quy m\u00f4 l\u1edbn. Tencent cho bi\u1ebft \u0111i\u1ec1u n\u00e0y \u0111\u1ebfn t\u1eeb vi\u1ec7c t\u00e1i c\u1ea5u tr\u00fac h\u1ea1 t\u1ea7ng v\u00e0o th\u00e1ng 2, do tr\u01b0\u1edfng b\u1ed9 ph\u1eadn AI Yao Shunyu d\u1eabn d\u1eaft, bao g\u1ed3m vi\u1ec7c x\u00e2y d\u1ef1ng l\u1ea1i to\u00e0n b\u1ed9 h\u1ec7 th\u1ed1ng pretraining v\u00e0 reinforcement learning.<\/p>\n

    C\u00e1ch ti\u1ebfp c\u1eadn n\u00e0y kh\u00e1c bi\u1ec7t \u0111\u00e1ng k\u1ec3 so v\u1edbi c\u00e1c lab AI Trung Qu\u1ed1c m\u1ed9t n\u0103m tr\u01b0\u1edbc, khi DeepSeek g\u00e2y b\u1ea5t ng\u1edd v\u1edbi model R1 nh\u1edd hi\u1ec7u qu\u1ea3 chi ph\u00ed.<\/p>\n

    D\u00f9 Hy3 v\u1eabn ch\u01b0a v\u01b0\u1ee3t qua c\u00e1c model flagship c\u1ee7a OpenAI v\u00e0 Google DeepMind, nh\u01b0ng x\u00e9t theo t\u1ef7 l\u1ec7 hi\u1ec7u su\u1ea5t tr\u00ean quy m\u00f4, Hy3 preview r\u1ea5t kh\u00f3 b\u1ecb b\u1ecf qua. Tr\u00ean b\u1ea3ng t\u1ed5ng h\u1ee3p benchmark v\u1ec1 agent, model n\u00e0y n\u1eb1m trong \u201cv\u00f9ng t\u1ed1i \u01b0u\u201d v\u1edbi ~295 t\u1ef7 tham s\u1ed1, v\u01b0\u1ee3t DeepSeek-V3.2 (600+ t\u1ef7 tham s\u1ed1) v\u00e0 ngang v\u1edbi Kimi-K2.5 (h\u01a1n 1 ngh\u00ecn t\u1ef7 tham s\u1ed1) nh\u01b0ng ch\u1ec9 c\u1ea7n chi ph\u00ed t\u00ednh to\u00e1n th\u1ea5p h\u01a1n \u0111\u00e1ng k\u1ec3.<\/p>\n

    \n
    \"\"<\/figure>\n<\/div>\n

    C\u00e1c model Hunyuan \u0111\u00e3 \u0111\u01b0\u1ee3c tri\u1ec3n khai tr\u00ean c\u00e1c n\u1ec1n t\u1ea3ng nh\u01b0 Yuanbao, CodeBuddy, WorkBuddy, QQ v\u00e0 Tencent Docs. Tr\u00ean CodeBuddy v\u00e0 WorkBuddy, \u0111\u1ed9 tr\u1ec5 token \u0111\u1ea7u ti\u00ean gi\u1ea3m 54%, th\u1eddi gian sinh n\u1ed9i dung end-to-end gi\u1ea3m 47%, v\u00e0 model c\u00f3 th\u1ec3 th\u1ef1c thi c\u00e1c workflow agent l\u00ean \u0111\u1ebfn 495 b\u01b0\u1edbc.<\/p>\n

    Tencent Cloud cung c\u1ea5p API v\u1edbi m\u1ee9c gi\u00e1 kho\u1ea3ng $0,18 cho m\u1ed7i 1 tri\u1ec7u input tokens v\u00e0 $0,59 cho m\u1ed7i 1 tri\u1ec7u output tokens, trong khi c\u00e1c g\u00f3i Token Plan c\u00e1 nh\u00e2n b\u1eaft \u0111\u1ea7u t\u1eeb kho\u1ea3ng $4,10\/th\u00e1ng.<\/p>\n

    \n

    B\u1ea3n tin Daily Debrief<\/strong>
    \nB\u1eaft \u0111\u1ea7u m\u1ed7i ng\u00e0y v\u1edbi nh\u1eefng tin t\u1ee9c n\u1ed5i b\u1eadt nh\u1ea5t, c\u00f9ng n\u1ed9i dung \u0111\u1ed9c quy\u1ec1n, podcast, video v\u00e0 nhi\u1ec1u h\u01a1n n\u1eefa.<\/p>","protected":false},"excerpt":{"rendered":"

    T\u00f3m t\u1eaft nhanh (In brief) Tencent Hy3 preview l\u00e0 m\u00f4 h\u00ecnh Mixture-of-Experts v\u1edbi 295 t\u1ef7 tham s\u1ed1, nh\u01b0ng ch\u1ec9 21 […]<\/p>","protected":false},"author":5,"featured_media":75811,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[220],"tags":[],"class_list":["post-75810","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tien-dien-tu"],"acf":[],"_links":{"self":[{"href":"https:\/\/hbbgroup.net\/vi\/wp-json\/wp\/v2\/posts\/75810","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/hbbgroup.net\/vi\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/hbbgroup.net\/vi\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/hbbgroup.net\/vi\/wp-json\/wp\/v2\/users\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/hbbgroup.net\/vi\/wp-json\/wp\/v2\/comments?post=75810"}],"version-history":[{"count":1,"href":"https:\/\/hbbgroup.net\/vi\/wp-json\/wp\/v2\/posts\/75810\/revisions"}],"predecessor-version":[{"id":76003,"href":"https:\/\/hbbgroup.net\/vi\/wp-json\/wp\/v2\/posts\/75810\/revisions\/76003"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/hbbgroup.net\/vi\/wp-json\/wp\/v2\/media\/75811"}],"wp:attachment":[{"href":"https:\/\/hbbgroup.net\/vi\/wp-json\/wp\/v2\/media?parent=75810"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/hbbgroup.net\/vi\/wp-json\/wp\/v2\/categories?post=75810"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/hbbgroup.net\/vi\/wp-json\/wp\/v2\/tags?post=75810"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}