{"id":142,"date":"2026-06-25T07:25:42","date_gmt":"2026-06-24T23:25:42","guid":{"rendered":"http:\/\/tantan.ink\/?p=142"},"modified":"2026-07-07T18:17:27","modified_gmt":"2026-07-07T10:17:27","slug":"minimax-m3","status":"publish","type":"post","link":"https:\/\/www.tantan.ink\/index.php\/2026\/06\/25\/minimax-m3\/","title":{"rendered":"MiniMax M3 \u53d1\u5e03\uff1a\u9996\u4e2a\u540c\u65f6\u505a\u5230\u7f16\u7801\u5f00\u6e90\u6700\u5f3a\u3001\u767e\u4e07\u4e0a\u4e0b\u6587\u3001\u539f\u751f\u591a\u6a21\u6001\u7684\u5f00\u653e\u6743\u91cd\u6a21\u578b"},"content":{"rendered":"

6 \u6708 1 \u65e5\uff0cMiniMax \u6b63\u5f0f\u53d1\u5e03\u4e86 M3\uff0c\u4e00\u4e2a\u5728\u597d\u51e0\u4e2a\u7ef4\u5ea6\u4e0a\u90fd\u8ba9\u4eba\u773c\u524d\u4e00\u4eae\u7684\u5f00\u653e\u6743\u91cd\u6a21\u578b\u3002\u5728\u672c\u5730 AI \u6a21\u578b\u5708\u5b50\u91cc\uff0c\u6211\u4eec\u89c1\u8fc7\u64c5\u957f\u7f16\u7801\u7684\u3001\u89c1\u8fc7\u5806\u957f\u4e0a\u4e0b\u6587\u7684\u3001\u4e5f\u89c1\u8fc7\u4e3b\u6253\u591a\u6a21\u6001\u7684\uff0c\u4f46 M3 \u662f\u7b2c\u4e00\u4e2a\u540c\u65f6\u505a\u5230\u8fd9\u4e09\u4ef6\u4e8b<\/strong>\u7684\u5f00\u6e90\u6743\u91cd\u6a21\u578b\u2014\u2014\u800c\u4e14\u6bcf\u4e00\u4ef6\u505a\u5f97\u90fd\u4e0d\u542b\u7cca\u3002<\/p>\n

\u8fd9\u4e0d\u662f\u4e00\u7bc7\u65b0\u95fb\u6458\u8981\uff0c\u6211\u66f4\u60f3\u4ece\u300c\u8fd9\u6a21\u578b\u5bf9\u672c\u5730 AI \u73a9\u5bb6\u610f\u5473\u7740\u4ec0\u4e48\u300d\u7684\u89d2\u5ea6\u6765\u804a\u804a\u3002<\/p>\n

\"MiniMax
MiniMax M3 \u5b98\u65b9\u53d1\u5e03\u6d77\u62a5\uff08\u6765\u6e90\uff1aminimax.io\uff09<\/em><\/p>\n

MSA\uff1a\u8ba9\u767e\u4e07\u4e0a\u4e0b\u6587\u4e0d\u518d\u53ea\u662f\u5671\u5934<\/h2>\n

\u5148\u8bf4\u67b6\u6784\u3002M3 \u7684\u6838\u5fc3\u521b\u65b0\u53eb MSA\uff08MiniMax Sparse Attention\uff09\uff0c\u672c\u8d28\u4e0a\u662f\u4e00\u79cd\u66f4\u806a\u660e\u7684\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236\u3002<\/p>\n

\u4e4b\u524d\u5927\u5bb6\u505a\u957f\u4e0a\u4e0b\u6587\uff0c\u8981\u4e48\u7528 DSA\uff08DeepSeek \u7684\u65b9\u6848\uff09\uff0c\u8981\u4e48\u7528 MoBA\uff08MoE-style attention\uff09\u3002MSA \u7684\u601d\u8def\u53c8\u8fdb\u4e86\u4e00\u6b65\uff1a\u5148\u628a KV cache \u5207\u6210\u5757\uff0c\u52a0\u4e00\u9053\u9884\u8fc7\u6ee4\uff0c\u53ea\u6311\u51fa\u8ddf\u5f53\u524d query \u76f8\u5173\u7684\u5757\u6765\u505a\u5b8c\u6574\u8ba1\u7b97\u3002\u542c\u8d77\u6765\u601d\u8def\u4e0d\u590d\u6742\uff0c\u4f46\u5173\u952e\u662f\u628a\u8fd9\u4e2a\u601d\u8def\u4e0a GPU \u505a\u5230\u4e86\u6781\u81f4\u2014\u2014\u540c\u4e00\u67e5\u8be2\u9700\u8981\u7684 KV \u5757\u8fde\u7eed\u8bfb\u53d6\u3001\u4e0d\u540c\u67e5\u8be2\u5171\u4eab\u540c\u4e00\u5757\u65f6\u4e00\u8d77 batch \u5904\u7406\uff0c\u6bcf\u5757\u53ea\u8bfb\u4e00\u6b21\u3002<\/p>\n

\u7ed3\u679c\u662f\u5b9e\u6253\u5b9e\u7684\uff1a\u5728 1M \u4e0a\u4e0b\u6587\u4e0b\uff0c\u6bcf\u4e2a token \u7684\u8ba1\u7b97\u91cf\u53ea\u6709\u524d\u4ee3\u7684 1\/20<\/strong>\uff0cprefill \u5feb\u4e86 9 \u500d<\/strong>\uff0cdecoding \u5feb\u4e86 15 \u500d<\/strong>\u3002\u5bf9\u6bd4\u73b0\u6709\u7684\u5f00\u6e90\u7a00\u758f\u6ce8\u610f\u529b\u65b9\u6848\uff0cMSA \u6574\u4f53\u4e5f\u5feb\u4e86 4 \u500d\u4ee5\u4e0a<\/strong>\u3002<\/p>\n

\u8fd9\u610f\u5473\u7740\u4ec0\u4e48\uff1f\u767e\u4e07\u7ea7\u4e0a\u4e0b\u6587\u7b2c\u4e00\u6b21\u5728\u5f00\u653e\u6743\u91cd\u6a21\u578b\u4e0a\u53d8\u5f97\u5b9e\u7528<\/strong>\uff0c\u800c\u4e0d\u662f\u505c\u7559\u5728\u8dd1\u5206\u6f14\u793a\u4e0a\u3002<\/p>\n

\"MSA\uff08MiniMax
MSA \u7a00\u758f\u6ce8\u610f\u529b\u67b6\u6784\u793a\u610f\u56fe\u2014\u2014\u5c06 KV cache \u5206\u5757\u9884\u8fc7\u6ee4\uff0c\u4ec5\u8ba1\u7b97\u76f8\u5173\u5757\uff08\u6765\u6e90\uff1aMiniMax \u5b98\u65b9\u535a\u6587\uff09<\/em><\/p>\n

\u8dd1\u5206\u4e0d\u662f\u91cd\u70b9\uff0c\u4f46\u786e\u5b9e\u80fd\u6253<\/h2>\n

SWE-Bench Pro 59.0%\u3001Terminal-Bench 2.1 66.0%\u3001BrowseComp 83.5\uff08\u8d85\u8fc7 Opus 4.7 \u7684 79.3\uff09\u2014\u2014\u5982\u679c\u53ea\u770b\u8fd9\u4e9b\u6570\u5b57\uff0cM3 \u5728\u7f16\u7801\u548c Agent \u7c7b\u4efb\u52a1\u4e0a\u5df2\u7ecf\u8fdb\u5165\u4e86\u95ed\u6e90\u524d\u6cbf\u6a21\u578b\u7684\u68af\u961f\u3002<\/p>\n

\u4e0d\u8fc7\u5728\u6211\u770b\u6765\uff0cMiniMax \u653e\u51fa\u7684\u4e09\u4e2a\u300c\u771f\u5b9e\u4efb\u52a1\u6f14\u793a\u300d\u66f4\u6709\u8bf4\u670d\u529b\uff1a<\/p>\n

\u8bba\u6587\u590d\u73b0<\/strong>\uff1aM3 \u82b1\u4e86 12 \u5c0f\u65f6\uff0c\u81ea\u4e3b\u590d\u73b0\u4e86\u4e00\u7bc7 ICLR 2025 \u7684\u8bba\u6587\u300aLearning Dynamics of LLM Finetuning\u300b\uff0c\u5b8c\u6210\u4e86 18 \u6b21\u4ee3\u7801\u63d0\u4ea4\u3001\u751f\u6210 23 \u5f20\u5b9e\u9a8c\u56fe\uff0c\u6700\u7ec8\u786e\u8ba4\u4e86 SFT \u8d8b\u52bf\u3001DPO \u6324\u538b\u6548\u5e94\u7b49\u6838\u5fc3\u53d1\u73b0\u3002\u5168\u7a0b\u65e0\u4eba\u5e72\u9884\u3002<\/p>\n

CUDA \u4f18\u5316<\/strong>\uff1a\u7ed9\u4e86\u4e2a\u4e0d\u5b8c\u6574\u7684 Triton \u9aa8\u67b6\u4ee3\u7801\uff0c\u8ba9 M3 \u4f18\u5316 Hopper FP8 GEMM \u5185\u6838\u3002\u5b83\u8dd1\u4e86\u5c06\u8fd1 24 \u5c0f\u65f6\u3001\u63d0\u4ea4\u4e86 147 \u6b21\uff0c\u628a\u786c\u4ef6\u5229\u7528\u7387\u4ece 7.6% \u4e00\u8def\u62c9\u5230\u4e86 71.3%<\/strong>\uff0c\u5b9e\u73b0\u4e86 9.4 \u500d\u7684\u52a0\u901f<\/strong>\u3002\u6700\u8ba9\u6211\u610f\u5916\u7684\u662f\uff0c\u6700\u4f18\u89e3\u51fa\u73b0\u5728\u7b2c 145 \u6b21\u63d0\u4ea4\u2014\u2014\u6362\u53e5\u8bdd\u8bf4\uff0c\u6a21\u578b\u5728\u7ecf\u5386\u591a\u6b21\u6027\u80fd\u5e73\u53f0\u671f\u540e\u4ecd\u7136\u6ca1\u6709\u653e\u5f03\uff0c\u800c\u662f\u5728\u6301\u7eed\u5c1d\u8bd5\u4e2d\u627e\u5230\u7a81\u7834\u53e3\u3002\u8fd9\u4e2a\u7279\u6027\u5728\u5b9e\u9645\u5f00\u53d1\u573a\u666f\u4e2d\u6bd4\u5355\u6b21\u8dd1\u5206\u66f4\u6709\u610f\u4e49\u3002<\/p>\n

PostTrainBench<\/strong>\uff1a\u7ed9 4 \u4e2a\u6ca1\u6709\u4e0b\u6e38\u80fd\u529b\u7684 Base \u6a21\u578b\uff0c\u8ba9 M3 \u81ea\u4e3b\u5b8c\u6210\u6570\u636e\u5408\u6210\u3001\u8bad\u7ec3\u3001\u8bc4\u4f30\u3001\u8fed\u4ee3\u300212 \u5c0f\u65f6\u540e\uff0c\u7ed3\u679c\u7565\u4f4e\u4e8e Opus 4.7 \u548c GPT-5.5\uff0c\u4f46\u660e\u663e\u9ad8\u4e8e\u5176\u4ed6\u6d4b\u8bd5\u7684\u6a21\u578b\u3002<\/p>\n

\u8bf4\u5b9e\u8bdd\uff0c\u8fd9\u4e09\u4e2a\u6f14\u793a\u52a0\u5728\u4e00\u8d77\uff0c\u6bd4\u4efb\u4f55\u5355\u4e00\u8dd1\u5206\u90fd\u66f4\u80fd\u8bf4\u660e\u95ee\u9898\uff1aM3 \u4e0d\u53ea\u662f\u300c\u80fd\u505a\u67d0\u4e00\u9053\u9898\u300d\uff0c\u800c\u662f\u80fd\u5728\u957f\u65f6\u95f4\u3001\u591a\u6b65\u9aa4\u7684\u771f\u5b9e\u4efb\u52a1\u4e2d\u6301\u7eed\u5de5\u4f5c<\/strong>\u2014\u2014\u8fd9\u5bf9 Agent \u573a\u666f\u6765\u8bf4\u662f\u5173\u952e\u80fd\u529b\u3002<\/p>\n

\"MiniMax
M3 \u5728\u7f16\u7801\u3001Agent\u3001\u641c\u7d22\u7b49\u8bc4\u6d4b\u4e2d\u7684\u8868\u73b0\uff08\u6765\u6e90\uff1aMiniMax \u5b98\u65b9\u535a\u6587\uff09<\/em><\/p>\n

\u539f\u751f\u591a\u6a21\u6001\uff0c\u800c\u4e0d\u662f\u540e\u88c5<\/h2>\n

M3 \u7684\u53e6\u4e00\u4e2a\u503c\u5f97\u5173\u6ce8\u7684\u7279\u5f81\u662f\uff1a\u5b83\u7684\u591a\u6a21\u6001\u662f\u4ece\u8bad\u7ec3\u7b2c\u4e00\u6b65\u5c31\u878d\u5165\u7684<\/strong>\uff0c\u4e0d\u662f\u62ff\u4e00\u4e2a\u7eaf\u6587\u672c\u6a21\u578b\u540e\u671f fine-tune \u56fe\u50cf\u80fd\u529b\u3002<\/p>\n

MiniMax \u91cd\u65b0\u8bbe\u8ba1\u4e86\u6574\u4e2a\u6587\u672c\u9884\u8bad\u7ec3\u7ba1\u7ebf\uff0c\u5927\u91cf\u4ea7\u51fa\u4e86\u6587\u672c+\u56fe\u50cf\u4ea4\u9519\u7684\u8bad\u7ec3\u6570\u636e\uff08\u89c4\u6a21\u7ea6 100T token\uff09\u3002\u8fd9\u79cd\u505a\u6cd5\u7684\u597d\u5904\u662f\uff0c\u6587\u672c\u548c\u56fe\u50cf\u7684\u8bed\u4e49\u7a7a\u95f4\u4ece\u4e00\u5f00\u59cb\u5c31\u878d\u5408\u5728\u4e00\u8d77\uff0c\u800c\u4e0d\u662f\u540e\u671f\u786c\u51d1\u3002<\/p>\n

\u8fd9\u4e00\u70b9\u5728\u8bba\u6587\u590d\u73b0\u4efb\u52a1\u4e2d\u4f53\u73b0\u5f97\u5f88\u6e05\u695a\uff1a\u6a21\u578b\u9700\u8981\u770b\u61c2\u8bba\u6587\u91cc\u7684\u516c\u5f0f\u3001\u56fe\u8868\u3001\u4ee5\u53ca\u4ee3\u7801\u7684\u5bf9\u5e94\u5173\u7cfb\uff0c\u8fd9\u6070\u6070\u662f\u300c\u539f\u751f\u591a\u6a21\u6001 + \u7f16\u7801\u80fd\u529b + \u957f\u4e0a\u4e0b\u6587\u300d\u4e09\u8005\u7ed3\u5408\u624d\u80fd\u5b8c\u6210\u7684\u3002<\/p>\n

\u5f00\u653e\u7a0b\u5ea6\u4e0e\u53ef\u7528\u6027<\/h2>\n

MiniMax \u627f\u8bfa\u5728\u53d1\u5e03\u540e 10 \u5929\u5185\u628a\u6743\u91cd\u548c\u6280\u672f\u62a5\u544a\u4e0a\u4f20 Hugging Face \u548c GitHub\u3002\u5230\u5199\u8fd9\u7bc7\u6587\u7ae0\u65f6\u5e94\u8be5\u5df2\u7ecf\u53ef\u4ee5\u4e0b\u8f7d\u4e86\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u914d\u5957\u7684 MiniMax Code<\/strong> Agent \u684c\u9762\u5e94\u7528\u4e5f\u4f1a\u5f00\u6e90\u2014\u2014\u8fd9\u6837\u4f60\u5728\u672c\u5730\u5c31\u80fd\u8dd1\u4e00\u4e2a\u80fd\u6301\u7eed\u5de5\u4f5c\u6570\u5929\u7684\u81ea\u4e3b\u7f16\u7801 Agent\u3002<\/p>\n

API \u5c42\u9762\uff0cToken \u8ba1\u5212\u4ece $20\/\u6708\uff08\u7ea6 1.7B tokens\uff09\u5230 $120\/\u6708\uff08\u7ea6 9.8B tokens\uff09\uff0c\u6240\u6709\u6a21\u6001\u5171\u7528\u989d\u5ea6\u3002\u652f\u6301 Thinking Mode \u5f00\u5173\uff0c512K \u4ee5\u5185\u4e0a\u4e0b\u6587\u6309\u6807\u51c6\u8d39\u7387\u8ba1\u8d39\u3002<\/p>\n

\"M3
MiniMax M3 \u4e0e DeepSeek-V4\u3001Qwen3 \u5728\u7f16\u7801\u3001\u957f\u4e0a\u4e0b\u6587\u3001\u591a\u6a21\u6001\u3001\u5f00\u653e\u6743\u91cd\u7b49\u7ef4\u5ea6\u5bf9\u6bd4<\/em><\/p>\n

\u600e\u4e48\u770b\u8fd9\u4ef6\u4e8b\uff1f<\/h2>\n

\u4f5c\u4e3a\u4e00\u4e2a\u504f\u5b9e\u5e72\u7684\u672c\u5730 AI \u7231\u597d\u8005\uff0c\u6211\u89c9\u5f97 M3 \u7684\u610f\u4e49\u4e0d\u5728\u4e8e\u300c\u53c8\u4e00\u4e2a\u6a21\u578b\u8dd1\u5206\u8d85\u8fc7 GPT\u300d\uff0c\u800c\u5728\u4e8e\uff1a<\/p>\n

    \n
  1. \u7a00\u758f\u6ce8\u610f\u529b\u7ec8\u4e8e\u6709\u4e86\u4e00\u4e2a\u5b9e\u7528\u3001\u53ef\u590d\u73b0\u7684\u5b9e\u73b0\u53c2\u8003<\/strong>\u3002MSA \u7684\u8bbe\u8ba1\u601d\u8def\u548c\u6027\u80fd\u6570\u636e\u5bf9\u793e\u533a\u540e\u7eed\u7684\u957f\u4e0a\u4e0b\u6587\u6a21\u578b\u5f88\u6709\u501f\u9274\u4ef7\u503c\u3002<\/li>\n
  2. \u5f00\u6e90\u6a21\u578b\u5728\u957f\u65f6\u95f4\u81ea\u4e3b\u4efb\u52a1\u4e0a\u7b2c\u4e00\u6b21\u62ff\u51fa\u4e86\u6709\u8bf4\u670d\u529b\u7684\u6f14\u793a<\/strong>\u3002\u6301\u7eed 12 \u5c0f\u65f6\u751a\u81f3 24 \u5c0f\u65f6\u4e0d\u95f4\u65ad\u5de5\u4f5c\uff0c\u8fd9\u5bf9\u73b0\u5b9e\u4e16\u754c\u4e2d\u7684\u5f00\u53d1\/\u7814\u7a76\u573a\u666f\u6765\u8bf4\u662f\u771f\u6b63\u7684\u751f\u4ea7\u529b\u3002<\/li>\n
  3. \u4e2d\u56fd AI \u516c\u53f8\u5728\u8fd9\u6ce2\u5f00\u6e90\u6d6a\u6f6e\u4e2d\u8d8a\u6765\u8d8a\u4e0d\u53ef\u5ffd\u89c6<\/strong>\u3002\u4ece DeepSeek \u5230 Qwen\uff0c\u518d\u5230\u73b0\u5728\u7684 MiniMax M3\uff0c\u5f00\u653e\u6743\u91cd\u6a21\u578b\u7684\u4e3b\u529b\u519b\u91cc\u4e2d\u56fd\u56e2\u961f\u7684\u6bd4\u4f8b\u5728\u660e\u663e\u4e0a\u5347\u3002<\/li>\n<\/ol>\n

    \u5982\u679c\u4f60\u5bf9 M3 \u611f\u5174\u8da3\uff0c\u73b0\u5728\u5c31\u53ef\u4ee5\u53bb Hugging Face \u641c MiniMax\/M3 \u62c9\u6743\u91cd\uff0c\u6216\u8005\u5148\u901a\u8fc7 API \u8bd5\u73a9\u3002\u6211\u4e2a\u4eba\u4f1a\u5173\u6ce8\u7684\u662f\u5b83\u5728 Mac \u4e0a\u7528 MLX \u8dd1\u7684\u6548\u679c\u2014\u2014\u6bd5\u7adf\u8fd9\u662f\u76ee\u524d\u6700\u80fd\u6253\u7684\u5f00\u6e90\u7f16\u7801\u6a21\u578b\u91cc\uff0c\u5c11\u6570\u627f\u8bfa\u80fd\u672c\u5730\u8dd1\u7684\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

    MiniMax \u53d1\u5e03 M3\u2014\u2014\u9996\u4e2a\u540c\u65f6\u5b9e\u73b0\u7f16\u7801 Agents \u524d\u6cbf\u30011M \u4e0a\u4e0b\u6587\u3001\u539f\u751f\u591a\u6a21\u6001\u7684\u5f00\u653e\u6743\u91cd\u6a21\u578b\u3002\u672c\u6587\u89e3\u6790 MSA \u67b6\u6784\u3001\u8dd1\u5206\u8868\u73b0\u4e0e\u771f\u5b9e\u4efb\u52a1\u6f14\u793a\u3002<\/p>\n","protected":false},"author":1,"featured_media":244,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2],"tags":[],"class_list":["post-142","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/142","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/comments?post=142"}],"version-history":[{"count":1,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/142\/revisions"}],"predecessor-version":[{"id":243,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/142\/revisions\/243"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/media\/244"}],"wp:attachment":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/media?parent=142"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/categories?post=142"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/tags?post=142"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}