
{"id":12611,"date":"2026-09-25T07:22:06","date_gmt":"2026-09-24T23:22:06","guid":{"rendered":"https:\/\/infernews.com\/blog\/bonsai-2-local-llm\/"},"modified":"2026-09-25T07:27:28","modified_gmt":"2026-09-24T23:27:28","slug":"bonsai-2-local-llm","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/bonsai-2-local-llm\/","title":{"rendered":"Bonsai 2 27B\uff1a5.9GB \u6a21\u578b\uff0c\u7b46\u96fb\u53ef\u672c\u5730\u8655\u7406\u8996\u89ba\u3001PDF \u8207\u5de5\u5177\u8abf\u7528"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/09\/pasted-138dfc27f3cb.jpg\" alt=\"Bonsai\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u60f3\u5728\u7b46\u8a18\u7c3f\u6216\u55ae\u5f35 GPU \u4e0a\u8655\u7406 27B \u7d1a\u6a21\u578b\uff0cBonsai 2 27B \u63d0\u4f9b\u4e86\u4e00\u689d\u66f4\u8cbc\u8fd1\u672c\u5730\u5de5\u4f5c\u7684\u8def\u7dda\u3002\u9019\u9805\u76ee\u5c6c\u672c\u5730\u904b\u884c\u7684\u4f4e\u4f4d\u5143\u591a\u6a21\u614b\u8a9e\u8a00\u6a21\u578b\u793a\u7bc4\u5de5\u5177\uff0c\u5be6\u969b\u8655\u7406\u63a8\u7406\u3001\u5716\u7247\u53ca PDF \u554f\u7b54\uff0c\u4ee5\u53ca\u9700\u8981\u5de5\u5177\u5f80\u8fd4\u7684\u4ee3\u7406\u5de5\u4f5c\u6d41\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Ternary Bonsai 2 27B \u5efa\u57fa\u65bc Qwen3.8 27B\uff0c\u4f7f\u7528 {-1, 0, +1} \u4e09\u503c\u6b0a\u91cd\u548c FP16 group-wise scaling\uff0c\u6a21\u578b\u5927\u5c0f\u7d04 5.9GB\uff0c\u9054\u5230 1.76 effective bits per weight\u3002\u9805\u76ee\u8072\u7a31\u4fdd\u7559\u5b8c\u6574\u7cbe\u5ea6\u6a21\u578b 98.2% \u7684\u6574\u9ad4\u57fa\u6e96\u8868\u73fe\uff0c\u6578\u5b78\u80fd\u529b\u53ea\u76f8\u5dee\u7d04\u534a\u5206\uff0c\u7de8\u7a0b\u8868\u73fe\u4ea6\u63a5\u8fd1\u57fa\u7dda\uff1b\u4ee3\u50f9\u662f\u5fc5\u9808\u4f7f\u7528 PrismML \u7684 llama.cpp binaries\uff0cstock llama.cpp \u7121\u6cd5\u76f4\u63a5\u8f09\u5165\u76f8\u95dc\u6a94\u6848\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u2022 262K-token context\uff0c\u914d\u5408 hybrid-attention backbone\uff0c\u8f03\u9069\u5408\u9577\u6587\u4ef6\u8655\u7406<br>\n\u2022 \u652f\u63f4\u7167\u7247\u3001\u622a\u5716\u53ca PDF \u554f\u7b54\uff0cllama.cpp \u8207 MLX \u5747\u53ef\u904b\u884c<br>\n\u2022 \u539f\u751f OpenAI-style tool_calls\uff0c\u4e26\u652f\u63f4 MCP servers<br>\n\u2022 PTQ1_0 \u7d04 1.75 bits per weight\uff1bPQ2_0 \u4ee5\u589e\u52a0\u7d04 1.3GB \u63db\u53d6\u8f03\u5feb\u63d0\u793a\u8655\u7406<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7531 setup \u6d41\u7a0b\u53d6\u5f97\u5c0d\u61c9\u6a21\u578b\u53ca\u81ea\u8a02\u5f8c\u7aef\uff0c\u518d\u900f\u904e\u793a\u7bc4\u4ecb\u9762\u958b\u555f\u804a\u5929\u3001\u8996\u89ba\u548c\u5de5\u5177\u529f\u80fd\u3002Mac \u53ef\u7528 Metal\uff0cLinux \u6216 Windows \u53ef\u9078 CUDA\u3001Vulkan\u3001ROCm\uff0c\u4ea6\u53ef\u7528 CPU\uff1b\u4ecb\u9762\u53ef\u8abf\u7bc0 reasoning effort\uff0c\u65b9\u4fbf\u6bd4\u8f03\u56de\u61c9\u901f\u5ea6\u8207\u63a8\u7406\u6df1\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9069\u5408\u9700\u8981\u96e2\u7dda\u8655\u7406\u6587\u4ef6\u3001\u60f3\u5728\u6709\u9650\u8a18\u61b6\u9ad4\u786c\u4ef6\u904b\u884c\u8f03\u5927\u6a21\u578b\uff0c\u6216\u8981\u628a\u672c\u5730\u6a21\u578b\u63a5\u5165 MCP \u5de5\u5177\u7684\u958b\u767c\u8005\u548c\u5c0f\u578b\u5718\u968a\u3002\u9805\u76ee\u63d0\u4f9b\u7684\u591a\u500b Bonsai\u3001Bonsai (1-bit) \u53ca Ternary-Bonsai \u578b\u865f\uff0c\u8b93\u786c\u4ef6\u5bb9\u91cf\u8207\u80fd\u529b\u4e4b\u9593\u6709\u66f4\u591a\u53d6\u6368\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/prismml.com\/news\/bonsai-2-27b\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/PrismML-Eng\/Bonsai-demo\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/huggingface.co\/collections\/prism-ml\/bonsai-2\" rel=\"noopener noreferrer\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Bonsai 2 27B \u4ee5\u8fd1\u4e4e\u5b8c\u6574\u7684\u63a8\u7406\u80fd\u529b\u7e2e\u81f3 5.9GB\uff0c\u517c\u9867\u8996\u89ba\u3001\u5de5\u5177\u8abf\u7528\u8207\u8d85\u9577\u4e0a\u4e0b\u6587\u3002<\/p>\n","protected":false},"author":8,"featured_media":12610,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,151,76,119,185,179,153,30,38,166,168,40,146,199],"tags":[],"class_list":["post-12611","post","type-post","status-publish","format-standard","hentry","category-133","category-mcp","category-76","category-119","category-qwen","category-nvidia","category-openai","category-image","category-38","category-mac","category-linux","category-llama","category-146","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12611","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=12611"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12611\/revisions"}],"predecessor-version":[{"id":12614,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12611\/revisions\/12614"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/12610"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=12611"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=12611"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=12611"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}