{"id":618,"date":"2026-08-02T06:58:27","date_gmt":"2026-08-01T22:58:27","guid":{"rendered":"http:\/\/www.tantan.ink\/index.php\/2026\/08\/02\/colibri-treats-744b-as-a-stream\/"},"modified":"2026-08-02T06:58:27","modified_gmt":"2026-08-01T22:58:27","slug":"colibri-treats-744b-as-a-stream","status":"publish","type":"post","link":"https:\/\/www.tantan.ink\/index.php\/2026\/08\/02\/colibri-treats-744b-as-a-stream\/","title":{"rendered":"colibr\u00ec\uff1a\u628a 744B \u6a21\u578b\u5f53\u6570\u636e\u6d41\u6d41\u5230 25GB \u7684\u7b14\u8bb0\u672c\u4e0a\u8dd1\u2014\u2014\u7eaf C \u5199\u7684 MoE \u63a8\u7406\u5f15\u64ce"},"content":{"rendered":"<p>\u770b\u5230 <strong>colibr\u00ec<\/strong>\uff08\u610f\u5927\u5229\u8bed\u7684&#8221;\u8702\u9e1f&#8221;\uff09\u8fd9\u4e2a\u540d\u5b57\u7684\u65f6\u5019\uff0c\u6211\u7b2c\u4e00\u53cd\u5e94\u662f\u53c8\u4e00\u4e2a\u5c0f\u5de7\u7684\u8f7b\u91cf\u63a8\u7406\u5de5\u5177\u3002\u70b9\u8fdb\u53bb\u624d\u53d1\u73b0\u8fd9\u4e2a\u9879\u76ee\u505a\u7684\u4e8b\u60c5\u5b8c\u5168\u4e0d\u662f\u90a3\u4e48\u56de\u4e8b\u2014\u2014\u5b83\u662f\u4e00\u4e2a<strong>\u7eaf C \u5199\u7684\u3001\u96f6\u4f9d\u8d56\u7684 MoE \u63a8\u7406\u5f15\u64ce<\/strong>\uff0c\u76ee\u6807\u662f\u5728 25 GB \u7684\u7b14\u8bb0\u672c RAM \u4e0a\u8dd1 <strong>744B \u53c2\u6570\u7684 GLM-5.2<\/strong>\u3002\u6574\u4e2a\u5f15\u64ce\u5355\u6587\u4ef6\uff0c<code>glm.c<\/code> \u52a0\u51e0\u4e2a\u5934\u6587\u4ef6\uff0c\u6ca1\u6709 BLAS\u3001\u6ca1\u6709 Python runtime\u3001\u6ca1\u6709 GPU \u4e5f\u80fd\u8dd1\u3002<\/p>\n<p>\u6211\u88c5\u4e86\u4e2a Mac mini M4\uff0816 GB\uff09\uff0c\u5e73\u65f6\u8dd1\u4e0d\u52a8\u4ec0\u4e48 70B \u4ee5\u4e0a\u7684\u6a21\u578b\uff0c\u5b98\u65b9\u6587\u6863\u91cc\u6070\u597d\u5217\u4e86 Metal \u540e\u7aef\u7684\u5b9e\u6d4b\u6570\u636e\u2014\u2014M4 Max \u662f 0.42 tok\/s\uff0cM5 Max \u52a0 47 GB \u5b66\u4e60\u7f13\u5b58\u80fd\u5230 2.06 tok\/s\u3002\u8fd9\u4ef6\u4e8b\u6700\u6253\u52a8\u6211\u7684\u662f\u5b83\u5904\u7406&#8221;<strong>\u8dd1\u4e0d\u52a8<\/strong>&#8220;\u8fd9\u4ef6\u4e8b\u7684\u601d\u8def\uff1a<strong>\u4e0d\u8981\u6c42\u6a21\u578b\u585e\u8fdb\u663e\u5b58\/\u5185\u5b58\uff0c\u800c\u662f\u628a\u6a21\u578b\u5f53\u6210\u8981\u6309\u9700\u642c\u5230\u8ba1\u7b97\u8fb9\u7684\u6570\u636e\u6d41<\/strong>\u3002\u8fd9\u6761\u601d\u8def\u53ef\u4ee5\u804a\u7684\u592a\u591a\u3002<\/p>\n<p><img decoding=\"async\" src=\"http:\/\/www.tantan.ink\/wp-content\/uploads\/2026\/08\/colibri-dashboard.png\" alt=\"dash\" class=\"alignnone size-full\" \/><\/p>\n<p><em>\u5b98\u65b9 Web \u63a7\u5236\u53f0\uff1a744B \u6a21\u578b\u5728 6\u00d7 RTX 5090 \u4e0a\u8dd1\uff0c4.0 tok\/s\u3001TTFT 1.6 s\u3001disk 0\uff08\u4e13\u5bb6\u5168\u9a7b\u7559\uff09<\/em><\/p>\n<h2>\u5b83\u7684\u6838\u5fc3\u5224\u65ad\uff1aMoE \u6a21\u578b\u4e0d\u9700\u8981&#8221;\u88c5\u5f97\u4e0b&#8221;<\/h2>\n<p>\u601d\u8def\u5176\u5b9e\u5f88\u76f4\u89c2\u3002\u4e00\u4e2a 744B \u7684 MoE \u6a21\u578b\uff08GLM-5.2 \u662f 75 \u5c42 MoE \u00d7 256 \u8def\u7531\u4e13\u5bb6 + MTP \u5934\uff09\uff0c\u6bcf\u4e2a token \u771f\u6b63\u6fc0\u6d3b\u7684\u53ea\u6709\u7ea6 40B \u53c2\u6570\uff0c\u800c\u771f\u6b63\u4f1a\u5728 token \u4e4b\u95f4\u6362\u8fdb\u6362\u51fa\u7684\u90e8\u5206\u53ea\u6709\u7ea6 11 GB\uff08\u8def\u7531\u4e13\u5bb6\uff09\u3002\u4e5f\u5c31\u662f\u8bf4 <strong>99% \u7684\u6743\u91cd\u5728\u5927\u90e8\u5206\u65f6\u95f4\u91cc\u662f&#8221;\u6b7b\u7684&#8221;<\/strong>\u3002<\/p>\n<p><img decoding=\"async\" src=\"http:\/\/www.tantan.ink\/wp-content\/uploads\/2026\/08\/colibri-sparse.png\" alt=\"sparse\" class=\"alignnone size-full\" \/><\/p>\n<p><em>\u6bcf token \u5b9e\u9645\u6fc0\u6d3b\u7684\u53ea\u6709 ~5.4% \u53c2\u6570<\/em><\/p>\n<p>colibr\u00ec \u5e72\u8106\u628a\u8fd9\u4e2a\u89c2\u5bdf\u505a\u6210\u4e86\u8bbe\u8ba1\uff1a<\/p>\n<ul>\n<li><strong>dense \u90e8\u5206<\/strong>\uff08attention + shared experts + embeddings\uff0c\u7ea6 17B \u53c2\u6570\uff09\u5e38\u9a7b RAM\uff0cint4 \u540e ~9.9 GB\uff1b<\/li>\n<li><strong>19,456 \u4e2a\u8def\u7531\u4e13\u5bb6<\/strong>\uff08\u6bcf\u4e2a ~19 MB int4\uff0c\u5171 ~370 GB\uff09\u4f4f\u5728 SSD \u4e0a\uff0c\u6309\u9700\u6d41\u5f0f\u8bfb\u53d6\uff1b<\/li>\n<li>\u5916\u52a0\u4e00\u5c42\u53ef\u9009\u7684 VRAM \u663e\u5b58 tier\uff0c\u7ed9\u6709 GPU \u7684\u4e3b\u673a\u5f53\u70ed\u7f13\u5b58\u3002<\/li>\n<\/ul>\n<p>\u5b98\u65b9\u628a\u8fd9\u5957\u673a\u5236\u7c7b\u6bd4\u6210 <strong>&#8220;\u6743\u91cd\u7684 JIT&#8221;<\/strong>\u2014\u2014\u4f20\u7edf\u7f16\u8bd1\u5668\u4e0d\u4f1a\u628a\u6574\u4e2a\u7a0b\u5e8f\u5168\u90e8\u7f16\u8bd1\u5b8c\uff0c\u5b83\u76ef\u7740\u70ed\u70b9\u8def\u5f84\uff0c&#8221;just in time&#8221;\u5730\u628a\u6b63\u5728\u8dd1\u7684\u4ee3\u7801\u7f16\u8bd1\u5230\u673a\u5668\u7801\u3002colibr\u00ec \u540c\u6837\u7684\u5224\u65ad\u4e0b\u5230\u53c2\u6570\u7a7a\u95f4\uff1a<strong>\u53c2\u6570\u4e0d\u662f\u5e38\u9a7b\u72b6\u6001\uff0c\u800c\u662f\u8981\u6309\u9700\u8de8 VRAM \/ RAM \/ NVMe \u4e09\u5c42\u7f13\u5b58\u5206\u9636\u6bb5\u642c\u8fd0\u7684\u6570\u636e<\/strong>\u3002\u8def\u7531\u662f\u53ef\u6d4b\u91cf\u7684\u7ed3\u6784\uff0c\u7ed3\u6784\u5c31\u53ef\u7f13\u5b58\u2014\u2014\u4f60\u8dd1\u5f97\u8d8a\u591a\uff0c\u6700\u5e38\u88ab\u8def\u7531\u5230\u7684\u90a3\u6279\u4e13\u5bb6\u5c31\u88ab\u4f60\u8dd1\u70ed\u4e86\u3002<\/p>\n<p><img decoding=\"async\" src=\"http:\/\/www.tantan.ink\/wp-content\/uploads\/2026\/08\/colibri-tiers.png\" alt=\"tiers\" class=\"alignnone size-full\" \/><\/p>\n<p><em>\u4e09\u5c42 memory hierarchy\uff1aVRAM \/ RAM \/ NVMe\uff0c\u540c\u4e00\u7ec4\u6743\u91cd\u53ea\u51b3\u5b9a\u901f\u5ea6\uff0c\u4e0d\u51b3\u5b9a\u8bed\u4e49<\/em><\/p>\n<h2>\u51e0\u4e2a\u5173\u952e\u5de5\u7a0b\u52a8\u4f5c<\/h2>\n<p>\u5149\u6709\u5206\u5c42\u4e0d\u591f\uff0c\u78c1\u76d8\u8bbf\u95ee\u7684\u5ef6\u8fdf\u662f\u51e0\u767e\u5fae\u79d2\u7ea7\u522b\uff0cCPU \u4e00\u7b97\u5c31\u662f\u51e0\u5341\u6beb\u79d2\uff0c\u5e38\u89c4\u505a\u6cd5\u7b49\u4e00\u6b21 I\/O \u5c31\u6162\u6b7b\u4e86\u3002colibr\u00ec \u7684\u51e0\u4e2a\u89e3\u6cd5\uff1a<\/p>\n<ul>\n<li><strong>Batch union<\/strong>\uff1a\u540c\u4e00\u6279 batch \u91cc\u53ea\u5bf9\u53bb\u91cd\u540e\u7684 expert \u8bfb\u4e00\u6b21\uff0c\u6bcf\u4e2a expert \u7684\u4e09\u4e2a\u77e9\u9635\uff08gate \/ up \/ down\uff09\u76f8\u90bb\u5b58\uff0c\u4e00\u6b21 <code>pread<\/code> \u8bfb\u5b8c\uff1b<\/li>\n<li><strong>PIPE \/ PILOT \u53cc\u6d41\u6c34\u7ebf<\/strong>\uff1a\u5f02\u6b65 I\/O \u6c60\u5728\u7b97\u7684\u540c\u65f6\u9884\u53d6\u4e0b\u4e00\u6279\u7f3a\u5931\u7684 expert\uff1b\u540c\u65f6\u8def\u7531\u662f\u53ef\u9884\u6d4b\u7684\uff0c\u5b98\u65b9\u5b9e\u6d4b&#8221;<strong>71.6% \u4e00\u5c42\u53ef\u9884\u6d4b<\/strong>&#8220;\uff0c\u6240\u4ee5\u53ef\u4ee5\u4e00\u5c42\u63d0\u524d\u9884\u53d6\uff1b<\/li>\n<li><strong>\u5b66\u4e60\u7f13\u5b58<\/strong>\uff1a\u6bcf\u4e2a turn \u628a\u8def\u7531\u5230\u7684 expert \u5199\u8fdb <code>.coli_usage<\/code>\uff0c\u5f15\u64ce\u81ea\u52a8 pin \u6700\u70ed\u7684\u51e0\u4e2a\uff0c\u8d8a\u7528\u8d8a\u5feb\uff1b<\/li>\n<li><strong>Dual-SSD<\/strong>\uff1aexpert \u53ea\u8bfb\uff0c\u6240\u4ee5\u628a\u6a21\u578b\u5b8c\u6574\u590d\u5236\u5230\u7b2c\u4e8c\u5757\u76d8\uff0c\u6309\u5e26\u5bbd\u6bd4\u54c8\u5e0c\u5206\u6d41\u5230\u4e24\u5757\u76d8\u4e0a\u5404\u8bfb\u5404\u7684\uff0c9 GB\/s + 3 GB\/s \u7684\u7ec4\u5408\u5b98\u65b9\u5b9e\u6d4b\u6bd4\u5355\u76d8\u5feb ~33%\u3002\u53ef\u4ee5&#8221;\u90e8\u5206\u955c\u50cf&#8221;\uff0c\u7b2c\u4e8c\u5757\u76d8\u653e\u4e0d\u4e0b\u5168\u6a21\u578b\u4e5f\u884c\u2014\u2014\u5b83\u4f1a\u6309\u5df2\u5b66\u5230\u7684\u8def\u7531\u5386\u53f2\u6311\u6700\u70ed\u7684 shard \u653e\u4e0a\u53bb\uff1b<\/li>\n<li><strong>KV \u538b\u7f29 57\u00d7<\/strong>\uff1aMLA attention \u7684 KV state \u6bcf token \u53ea\u7528 576 \u4e2a float\uff08\u4e0d\u662f 32768\uff09\uff0c\u5e76\u4e14\u5199\u5230 <code>.coli_kv<\/code>\uff0c\u5bf9\u8bdd\u91cd\u5f00\u96f6 prefill\uff0c\u9010\u5b57\u8282\u76f8\u540c\u4e8e\u4e0d\u4e2d\u65ad\u7684\u4f1a\u8bdd\u3002<\/li>\n<\/ul>\n<h2>\u5728\u6211\u673a\u5668\u4e0a\u80fd\u8dd1\u591a\u5feb\uff1f<\/h2>\n<p>\u5b98\u65b9\u53d1\u5e03\u7684\u540c\u5f15\u64ce\u3001\u540c int4 \u6a21\u578b\u5728\u4e0d\u540c\u786c\u4ef6\u4e0a\u7684\u5b9e\u6d4b\u901f\u5ea6\uff08greedy decode\uff09\uff1a<\/p>\n<p><img decoding=\"async\" src=\"http:\/\/www.tantan.ink\/wp-content\/uploads\/2026\/08\/colibri-ladder.png\" alt=\"ladder\" class=\"alignnone size-full\" \/><\/p>\n<p><em>\u540c\u4e00\u4e2a\u5f15\u64ce\uff0c\u540c\u4e00\u4e2a int4 \u6a21\u578b\u2014\u2014\u53ea\u6362\u786c\u4ef6\uff0c\u533a\u522b\u53ea\u662f\u4e13\u5bb6\u4f4f\u54ea\u91cc<\/em><\/p>\n<p>\u6362\u53e5\u8bdd\u8bf4\uff0c<strong>\u4e00\u5757 SSD + 25 GB RAM \u4e5f\u80fd\u8dd1\u5f97\u52a8 744B<\/strong>\uff0c\u53ea\u662f 0.05\u20130.1 tok\/s \u8fd9\u79cd&#8221;\u6253\u5b57\u5458\u90fd\u6bd4\u4f60\u5feb&#8221;\u7684\u6c34\u5e73\u3002Mac \u8fd9\u8fb9 Metal \u540e\u7aef\u662f\u5b9e\u9a8c\u6027\u7684\uff0c\u6587\u6863\u91cc\u7ed9\u7684\u6570\u636e\u662f\uff1a<\/p>\n<ul>\n<li><strong>M4 Max\uff08128 GB, \u70ed\u7f13\u5b58, MTP on\uff09<\/strong>\uff1aCPU 0.30 \u2192 Metal 0.42 tok\/s\uff08~1.4\u00d7\uff09\u3002\u6253\u5f00 <code>DIRECT=1<\/code> \u518d\u5feb\u4e00\u4e9b\uff0c\u5bf9\u6bd4\u8fd9\u53f0\u673a\u5668\u7684\u51b7\u542f\u52a8 <strong>~3\u00d7<\/strong>\u3002<\/li>\n<li><strong>M5 Max\uff08\u5e26 46.9 GB \u5b66\u4e60\u7f13\u5b58\uff09<\/strong>\uff1a<strong>2.06 tok\/s<\/strong>\u3002<\/li>\n<\/ul>\n<p>16 GB Mac \u60f3\u8dd1 int4 GLM-5.2 dense \u90e8\u5206\uff08~9.9 GB\uff09\u5c31\u88c5\u4e0d\u4e0b\uff0c\u6240\u4ee5\u5b98\u65b9\u6587\u6863\u5bf9 Apple Silicon \u8fd9\u4e00\u6863\u5199\u7684\u662f&#8221;experimental&#8221;\u3002\u4f46 7B \u7684 OLMoE\u3001Inkling \u5728\u66f4\u5c0f RAM \u4e0a\u6709\u4e13\u95e8\u7684 dense \u91cf\u5316\u5de5\u5177\u2014\u2014\u6587\u6863\u91cc\u660e\u786e\u5199\u4e86&#8221;Inkling \u5728 RAM \u7d27\u5f20\u7684\u4e3b\u673a\u4e0a\u8981\u5148\u628a dense \u90e8\u5206 int4 \u91cf\u5316\u5230 15.3 GB \u624d\u884c\uff0c\u4e14 trade-off \u90fd\u5199\u5728\u6587\u6863\u91cc&#8221;\u3002\u8fd9\u79cd\u628a\u5c40\u9650\u76f4\u63a5\u5199\u51fa\u6765\u3001\u4e0d\u505a\u6a21\u7cca\u627f\u8bfa\u7684\u6001\u5ea6\uff0c\u662f\u8fd9\u4e2a\u9879\u76ee\u8ba9\u6211\u89c9\u5f97\u5f88\u8212\u670d\u7684\u4e00\u70b9\u3002<\/p>\n<h2>\u6211\u770b\u5230\u7684\u51e0\u4e2a\u5751\uff08\u6ca1\u8e29\u5b9e\uff0c\u4f46\u6587\u6863\/issue \u5199\u5f97\u660e\u660e\u767d\u767d\uff09<\/h2>\n<ul>\n<li><strong>\u522b\u4e0b\u9519 int4 \u5bb9\u5668<\/strong>\u3002Hugging Face \u4e0a\u6709\u8001\u7248&#8221;per-row int4&#8243;\uff08<code>mateogrgic\/...<\/code>\u3001<code>jlnsrk\/...<\/code>\uff09\uff0c\u8d28\u91cf\u7ea6\u5dee 9pp\uff0c\u662f\u65e9\u671f think-mode \u5faa\u73af\u8f93\u51fa\u7684\u6839\u56e0\u3002\u8981\u7528\u5b98\u65b9\u7684 <strong>gs64 group-scaled<\/strong> + <strong>int8 MTP head<\/strong> \u7684\u90a3\u4e2a\u3002<\/li>\n<li><strong>MTP head \u5fc5\u987b\u662f int8<\/strong>\u3002int4 \u5934\u5728\u5b9e\u6d4b\u4e2d\u8349\u7a3f\u63a5\u53d7\u7387\u4f1a\u6389\u5230 0\u20134%\u3002\u68c0\u67e5\u65b9\u6cd5\uff1a<code>ls -l &lt;model&gt;\/out-mtp-*<\/code>\uff0cint8\uff08\u6b63\u786e\uff09\u7684\u6587\u4ef6\u5927\u5c0f\u662f 3527131672 \/ 5366238584 \/ 1065950496\u3002<\/li>\n<li><strong>MTP \u4e0d\u662f\u94f6\u5f39<\/strong>\u3002\u5728 expert hit rate 85% \u9644\u8fd1\uff0cMTP \u5b9e\u6d4b\u6709 ~32% \u7684 regression\uff0c\u6587\u6863\u660e\u786e\u5199\u4e86\u2014\u2014\u6295\u673a\u89e3\u7801\u5728\u4e0d\u70ed\u7684\u65f6\u5019\u662f\u4e8f\u7684\uff0c\u8981 <code>DRAFT=0<\/code> \u5173\u6389\u3002<\/li>\n<li><strong>DIRECT=1 \u4e0d\u4e00\u5b9a\u8d62<\/strong>\u3002<code>O_DIRECT<\/code> \u5728\u5e26 DRAM cache \u7684 NVMe \u4e0a\u662f\u5927\u8d62\uff08\u5b9e\u6d4b +34% decode\uff09\uff0c\u4f46\u5728 QLC\u3001DRAM-less\u3001\u865a\u62df\u76d8\u4e0a\u53ef\u80fd\u662f\u4e2d\u6027\u751a\u81f3\u8d1f\u503c\u3002\u8981\u5148\u8bd5\u518d\u7559\u3002<\/li>\n<li><strong>Metal \u540e\u7aef prefill \u7684\u6d6e\u70b9\u7cbe\u5ea6<\/strong>\u3002Decode \u662f byte-identical \u7684\uff0c\u4f46 prefill \u5927 GEMM \u5728 GPU \u4e0a\u8dd1\u4f1a\u6362\u7d2f\u52a0\u987a\u5e8f\uff0c\u5728 teacher-forced oracle \u6bd4\u8f83\u7684\u6781\u7aef case \u91cc\u53ef\u80fd\u6311\u4e0d\u540c\u7684 top token\u3002\u6587\u6863\u5efa\u8bae <code>COLI_METAL_GEMM_MIN=100000<\/code> \u628a\u6240\u6709 GEMM \u7559\u5728 CPU \u4e0a\u4fdd\u6301 bit-exact\u3002<\/li>\n<li><strong>\u955c\u50cf\u4e0d\u662f\u53cc\u5199<\/strong>\u3002<code>COLI_MODEL_MIRROR<\/code> \u542f\u52a8\u65f6\u6821\u9a8c\u5927\u5c0f\u548c safetensors header \u4e00\u81f4\uff1b\u4e2d\u9014\u62d4\u7b2c\u4e8c\u5757\u76d8\u4f1a\u56de\u9000\u5230\u4e3b\u76d8\uff08\u4e00\u6b21\u8b66\u544a\uff0c\u4e0d\u5d29\uff09\uff1b\u955c\u50cf\u6c38\u8fdc\u662f\u53ea\u8bfb\u7684\u3002<\/li>\n<li><strong>KV \u538b\u7f29\u4e0d\u5f71\u54cd forward<\/strong>\u3002forward pass \u7528 <code>transformers<\/code> oracle \u9a8c\u8bc1\uff08teacher-forcing 30\u201332\/32\uff09\uff0c\u91cf\u5316\u7248\u672c\u548c\u539f\u7248\u5728\u91c7\u6837\u8def\u5f84\u4e0a token-exact\u3002<\/li>\n<\/ul>\n<h2>\u8fd9\u9879\u76ee\u503c\u4e0d\u503c\u5f97\u88c5<\/h2>\n<p>\u5982\u679c\u4f60\u662f 16 GB Mac mini \u8fd9\u79cd\u914d\u7f6e\uff0c\u7b54\u6848\u662f<strong>\u76f4\u63a5\u88c5\u6a21\u578b\u7684\u4eba\u4e0d\u7528\u7ea0\u7ed3<\/strong>\u2014\u2014\u76ee\u524d OLMoE\uff087B \/ 1B \u6fc0\u6d3b\uff09\u624d\u662f\u8fd9\u4e00\u6863\u7684\u751c\u70b9\uff0c744B GLM-5.2 \u4f60\u8dd1\u4e0d\u5feb\u4e5f\u8dd1\u4e0d\u6ee1\u3002\u4f46\u4f60\u53ef\u4ee5\u88c5\u4e0a\u8fd9\u4e2a\u5f15\u64ce\uff0c\u770b\u5b83\u7684 Web \u63a7\u5236\u53f0 + Brain + Atlas \u4e09\u9875\u7684&#8221;<strong>\u4e13\u5bb6\u56fe\u8c31<\/strong>&#8220;\u2014\u2014<strong>Atlas \u628a 13260 \u4e2a\u88ab\u6d4b\u8fc7\u8def\u7531\u884c\u4e3a\u7684 expert \u753b\u6210 3D \u661f\u7cfb<\/strong>\uff0c\u4f4d\u7f6e\u662f\u5b9e\u6d4b\u7684\u8def\u7531\u4eb2\u548c\u5ea6\u3001\u4e0d\u662f\u5b66\u51fa\u6765\u7684 embedding\u3002\u9f20\u6807 hover \u80fd\u770b\u6bcf\u4e2a expert \u7684\u4e3b\u9898\u504f\u597d\u3002\u6211\u770b\u5230 poetry\u3001law\u3001Chinese\u3001SQL \u8fd9\u4e9b\u805a\u7c7b\u5df2\u7ecf\u81ea\u52a8\u6563\u5f00\u4e86\u3002\u8fd9\u662f\u8fd9\u4e2a\u9879\u76ee\u6700\u8ff7\u4eba\u3001\u6700\u4e0d\u50cf\u4f20\u7edf\u63a8\u7406\u5f15\u64ce\u7684\u4e00\u9762\u3002<\/p>\n<p><img decoding=\"async\" src=\"http:\/\/www.tantan.ink\/wp-content\/uploads\/2026\/08\/colibri-atlas.png\" alt=\"atlas\" class=\"alignnone size-full\" \/><\/p>\n<p><em>\u5b98\u65b9 Atlas \u9875\uff1a13,260 \u4e2a\u88ab\u6d4b\u8fc7\u8def\u7531\u884c\u4e3a\u7684\u4e13\u5bb6\uff0c\u6309\u5b9e\u6d4b\u8def\u7531\u4eb2\u548c\u5ea6\u6563\u6210 3D \u661f\u7cfb\uff0c\u80fd\u770b\u51fa poetry\u3001law\u3001Chinese\u3001SQL \u7b49\u4e3b\u9898\u805a\u7c7b<\/em><\/p>\n<p>\u5982\u679c\u4f60\u662f 128 GB \u5185\u5b58 + \u4e00\u5757 SSD \u7684\u684c\u9762\u7528\u6237\uff0c\u7b54\u6848\u662f<strong>\u5f88\u503c\u5f97<\/strong>\u2014\u20141.8 tok\/s \u7684 GLM-5.2 \u8dd1 chat \u662f\u80fd\u7528\u7684\uff0c\u800c\u4e14\u8fd9\u662f\u4e2a\u80fd\u8ba9\u4f60&#8221;\u7528\u7740\u8dd1&#8221;\u7684\u9879\u76ee\uff1a\u6bcf\u6761\u6d88\u606f\u90fd\u4f1a\u8ba9\u4f60\u66f4\u70ed\u7684\u90a3\u6279 expert \u8fdb RAM\uff0c\u8d8a\u804a\u8d8a\u5feb\u3002<\/p>\n<p>\u5982\u679c\u4f60\u662f\u591a GPU \u4e3b\u673a\u7528\u6237\uff0c<strong>v1.4.0 \u521a\u51fa\u7684 Vulkan \u540e\u7aef<\/strong>\u628a 7\u00d7 AMD RX 580 \u8fd9\u79cd\u5b98\u65b9\u5df2\u7ecf\u653e\u5f03\u7684\u5361\u90fd\u7eb3\u5165\u4e86\uff08ROCm \u4e0d\u652f\u6301\uff09\uff0c\u8fd8\u80fd\u5728 RDNA4 \u4e0a\u548c ROCm \u62fc\u4e00\u4e0b\u3002Vulkan \u4e0d\u7ed1\u9a71\u52a8\uff0c\u662f\u8fd9\u4e2a\u9879\u76ee\u7684\u53e6\u4e00\u4e2a\u6709\u610f\u601d\u7684\u8d4c\u6ce8\u3002<\/p>\n<p><img decoding=\"async\" src=\"http:\/\/www.tantan.ink\/wp-content\/uploads\/2026\/08\/colibri-brain.png\" alt=\"brain\" class=\"alignnone size-full\" \/><\/p>\n<p><em>\u5b98\u65b9 Brain \u9875\uff1a19,456 \u4e2a\u4e13\u5bb6\u5f53\u4f5c\u6210\u6d3b\u7684\u8111\u76ae\u5c42\uff0c\u989c\u8272\u662f\u5b58\u50a8 tier\u3001\u4eae\u5ea6\u662f\u8def\u7531\u70ed\u5ea6\uff0c\u6bcf\u4e00\u8f6e\u8def\u7531\u5230\u7684\u4e13\u5bb6\u4f1a\u95ea\u767d<\/em><\/p>\n<h2>\u5b89\u88c5\u5907\u5fd8<\/h2>\n<p>\u6b65\u9aa4\u6781\u7b80\uff1a<\/p>\n<pre><code># 1. \u4e0b\u8f7d\u5bf9\u5e94\u5e73\u53f0\u7684 release\uff0c\u89e3\u538b\nmkdir colibri &amp;&amp; tar xzf colibri-*.tar.gz -C colibri &amp;&amp; cd colibri\npython3 coli info\n\n# 2. \u4e0b\u6a21\u578b\uff08372 GB int4 gs64 + int8 MTP \u5934\uff09\n# https:\/\/huggingface.co\/mastouri\/GLM-5.2-colibri-int4-g64-with-int8-mtp\n\n# 3. \u8dd1\nCOLI_MODEL=\/nvme\/glm52_i4 .\/coli chat\nCOLI_MODEL=\/nvme\/glm52_i4 .\/coli web      # \u8d77 API + Web \u63a7\u5236\u53f0\n\n# Mac \u52a0 Metal \u540e\u7aef\ncd c &amp;&amp; make colibri METAL=1\nCOLI_METAL=1 COLI_MODEL=\/path\/glm52_i4 .\/coli chat --ram 96\n<\/code><\/pre>\n<p>License \u662f Apache 2.0\uff0c\u6a21\u578b\u6743\u91cd\u9075\u5faa\u5404\u81ea\u4e0a\u6e38\u7684\u534f\u8bae\uff08GLM-5.2 \u662f MIT\uff09\u3002<\/p>\n<h2>\u5c0f\u7ed3<\/h2>\n<p>\u6211\u5bf9\u8fd9\u4e2a\u9879\u76ee\u7684\u6001\u5ea6\u6bd4\u9884\u60f3\u66f4\u6b63\u9762\u3002\u4e09\u4ef6\u4e8b\u505a\u5f97\u5bf9\uff1a<\/p>\n<ul>\n<li><strong>\u628a\u5047\u8bbe\u6807\u5f97\u6e05\u6e05\u695a\u695a<\/strong>\u3002\u6bcf\u6761\u4f18\u5316\u90fd\u9644\u4e86\u8bc1\u636e + &#8220;this is what we still need to test&#8221;\uff0c\u8fde MTP \u90a3\u4e2a 32% regression \u90fd\u5199\u5728\u663e\u773c\u4f4d\u7f6e\uff1b<\/li>\n<li><strong>\u628a\u5206\u9636\u6bb5\u642c\u8fd0\u7684\u601d\u8def\u8bb2\u900f\u4e86<\/strong>\u3002\u4ece dual SSD \u5230 partial mirror\uff0c\u4ece RAM \u7d27\u5f20\u65f6\u964d\u5bc6\u5230 KV \u538b\u7f29\uff0c\u90fd\u662f&#8221;\u7ea6\u675f\u5148\u627f\u8ba4\uff0c\u518d\u60f3\u529e\u6cd5\u7ed5&#8221;\uff1b<\/li>\n<li><strong>\u628a\u5de5\u5177\u505a\u6210\u4e86\u7814\u7a76\u5e73\u53f0<\/strong>\u3002Atlas \u90a3\u4e00\u9875\u4e0d\u662f\u6f02\u4eae\u5c31\u5b8c\u4e8b\u2014\u2014\u5b83\u662f\u8def\u7531\u7ed3\u6784\u6709\u53ef\u6d4b\u89c4\u5f8b\u7684\u5b9e\u8bc1\uff0c&#8221;structure is cacheable&#8221;\u8fd9\u4e2a\u5224\u65ad\u4e0d\u80fd\u7a7a\u53e3\u8bf4\uff0c\u5f97\u753b\u51fa\u6765\u7ed9\u8bfb\u8005\u770b\u3002<\/li>\n<\/ul>\n<p>2026 \u5e74\u5f00\u6e90\u5927\u6a21\u578b\u771f\u6b63\u8fdb\u5165&#8221;<strong>\u672c\u5730\u63a8\u7406<\/strong>&#8220;\u8fd9\u4e2a\u8bdd\u9898\u65f6\uff0c\u786c\u4ef6\u95e8\u69db\u5df2\u7ecf\u4e0d\u662f&#8221;\u4f60\u6709\u591a\u5c11\u663e\u5b58&#8221;\uff0c\u800c\u662f&#8221;\u4f60\u613f\u4e0d\u613f\u610f\u628a\u5b58\u50a8 \/ \u5185\u5b58 \/ \u663e\u5b58\u5f53\u4f5c\u4e00\u4e2a\u5206\u5c42\u7684\u6570\u636e\u6d41&#8221;\u3002colibr\u00ec \u8fd9\u4e2a\u9879\u76ee\u628a\u8fd9\u4e2a\u601d\u8def\u6267\u884c\u5230\u4e86\u6211\u76ee\u524d\u770b\u5230\u7684\u6700\u786c\u6838\u7684\u7a0b\u5ea6\u2014\u2014\u800c\u4e14\u5b83\u4eca\u5929\u5c31\u80fd\u88c5\u3001\u4eca\u5929\u5c31\u80fd\u8dd1\u3002<\/p>\n<p>\u2014\u2014 \u7b49\u6211\u90a3\u5757 SSD \u5230\u8d27\u4e86\u518d\u5b9e\u6d4b\u4e00\u6b21 M4 Mac \u4e0a\u7684 Metal decode\u3002<\/p>\n<p>\u4ed3\u5e93\uff1a<a href=\"https:\/\/github.com\/JustVugg\/colibri\" target=\"_blank\" rel=\"noopener\">github.com\/JustVugg\/colibri<\/a> \uff5c \u6a21\u578b\uff1a<a href=\"https:\/\/huggingface.co\/mastouri\/GLM-5.2-colibri-int4-g64-with-int8-mtp\" target=\"_blank\" rel=\"noopener\">huggingface.co\/mastouri\/GLM-5.2-colibri-int4-g64-with-int8-mtp<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>colibr\u00ec \u662f\u4e2a\u7eaf C \u5199\u7684\u96f6\u4f9d\u8d56 MoE \u63a8\u7406\u5f15\u64ce\uff0c\u628a 744B \u7684 GLM-5.2 \u6a21\u578b\u5f53\u6210\u6309\u9700\u8de8 VRAM \/ RAM \/ NVMe \u4e09\u5c42\u7f13\u5b58\u5206\u9636\u6bb5\u642c\u8fd0\u7684\u6570\u636e\u6d41\uff0c\u7528\u300c\u6743\u91cd\u7684 JIT\u300d\u7684\u601d\u8def\u5728 25 GB \u7b14\u8bb0\u672c RAM \u4e0a\u8dd1\u5f97\u52a8\u3002<\/p>\n","protected":false},"author":2,"featured_media":612,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[2,10],"tags":[235,237,239,236,238],"class_list":["post-618","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-software","tag-colibri","tag-glm","tag-llm","tag-moe","tag-238"],"_links":{"self":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/618","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\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/comments?post=618"}],"version-history":[{"count":0,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/posts\/618\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/media\/612"}],"wp:attachment":[{"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/media?parent=618"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/categories?post=618"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.tantan.ink\/index.php\/wp-json\/wp\/v2\/tags?post=618"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}