
{"id":12125,"date":"2026-09-11T00:07:30","date_gmt":"2026-09-10T16:07:30","guid":{"rendered":"https:\/\/infernews.com\/blog\/edge0-moe\/"},"modified":"2026-09-11T00:24:13","modified_gmt":"2026-09-10T16:24:13","slug":"edge0-moe","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/edge0-moe\/","title":{"rendered":"Edge0 \u958b\u6e90\u6846\u67b6\uff1aMac \u672c\u6a5f\u63a8\u8ad6 35B MoE\uff0c\u8a18\u61b6\u9ad4\u50c5\u4f54\u7528 2.9 GB"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/09\/pasted-22491fb5911f.jpg\" alt=\"edge0\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">MoE \u6a21\u578b\u53c3\u6578\u898f\u6a21\u6108\u63a8\u6108\u5927\uff0c\u4f46\u771f\u6b63\u80fd\u5728\u6d88\u8cbb\u7d1a\u786c\u9ad4\u672c\u6a5f\u8dd1\u5f97\u52d5\u7684\u65b9\u6848\u4e00\u76f4\u4e0d\u591a\u3002Edge0 \u60f3\u8655\u7406\u7684\u5c31\u662f\u9019\u500b\u843d\u5dee\uff1a\u5b83\u628a SSD \u5c08\u5bb6\u5378\u8f09\u3001Recover-LoRA \u9084\u539f\u3001\u4ee5\u53ca prerouter \u8def\u7531\u9810\u6e2c\u9019\u4e09\u62db\u7d44\u5408\u6210\u4e00\u5957\u53ef\u64f4\u5c55\u6846\u67b6\uff0c\u76ee\u524d\u4ee5 MLX \u5f8c\u7aef\u8dd1\u5728 Apple Silicon\uff08M1 \u81f3 M4\uff09\u4e0a\uff0cCUDA \u5f8c\u7aef\u5217\u5728\u8def\u7dda\u5716\u4f46\u5c1a\u672a\u652f\u63f4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u96a8\u6846\u67b6\u91cb\u51fa\u5169\u500b\u6a21\u578b\u6a94\uff1aedge0-35b\uff08\u57fa\u65bc Qwen3.5-MoE 35B-A3B\uff09\u8207 edge0-8b\uff08\u57fa\u65bc Ling 3.0 bailing \u6df7\u5408\u67b6\u69cb\uff09\uff0c\u7686\u4ee5 4-bit \u91cf\u5316\u91cb\u51fa\u3002LoRA \u914d\u91cd\u8207 prerouter \u982d\u5df2\u96a8 checkpoint \u4e00\u540c\u767c\u4f48\uff0c<code>edge0 serve <\/code>\u5373\u53ef\u76f4\u63a5\u8f09\u5165\u8a13\u7df4\u597d\u7684\u5b8c\u6574\u7ba1\u7dda\uff0c\u4e0d\u7528\u984d\u5916\u62fc\u6e4a\u7d44\u4ef6\u3002\u77ed\u4e0a\u4e0b\u6587\u60c5\u5883\u4e0b\uff0cedge0-35b \u5cf0\u503c\u6d3b\u52d5\u8a18\u61b6\u9ad4\u7d04 2.9 GB\uff0cedge0-8b \u7d04 1.0 GB\uff1b\u5c08\u5bb6\u6b0a\u91cd\u4ee5 mmap \u65b9\u5f0f\u6309\u9700\u8b80\u53d6\uff0c\u4e0d\u6703\u9810\u5148\u4f54\u7528 RAM\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>Tier<\/th><th>Released checkpoint<\/th><th>Inference profile<\/th><\/tr><\/thead><tbody><tr><td><code>edge0-35b<\/code><\/td><td><a href=\"https:\/\/huggingface.co\/Edge0\/Edge0-35B-A3B-preview\"><code>Edge0\/Edge0-35B-A3B-preview<\/code><\/a><\/td><td>4-bit, 40 layers, 256 experts, prerouter K=4<\/td><\/tr><tr><td><code>edge0-8b<\/code><\/td><td><a href=\"https:\/\/huggingface.co\/Edge0\/Edge0-8B-A1B-preview\"><code>Edge0\/Edge0-8B-A1B-preview<\/code><\/a><\/td><td>4-bit, 24 layers, 128 experts, prerouter K=8<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f7f\u7528\u4ecb\u9762\u63d0\u4f9b AutoModel\u3001AutoConfig\u3001AutoEngine \u81ea\u52d5\u6309\u6a21\u578b\u540d\u7a31\u89e3\u6790\u5c64\u7d1a\u3002\u6240\u6709 MLX \u7a0b\u5f0f\u78bc\u96c6\u4e2d\u5728 <code>edge0\/backends\/mlx\/<\/code>\uff0c\u6838\u5fc3\u908f\u8f2f\u53ea\u4f9d\u8cf4\u62bd\u8c61\u5c64\uff0c\u56e0\u6b64\u65e5\u5f8c\u52a0\u5165 CUDA \u7b49\u5f8c\u7aef\u6642\u4e0d\u5fc5\u91cd\u5beb\u8abf\u5ea6\u908f\u8f2f\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u985e\u578b<\/strong>\uff1a\u958b\u6e90 MoE \u63a8\u8ad6\u6846\u67b6\uff08\u642d\u914d\u5169\u500b\u9810\u8a13\u7df4 checkpoint\uff09\u3002<\/li>\n\n\n\n<li><strong>\u8cc7\u6e90\u9580\u6abb<\/strong>\uff1a4-bit 35B \u6a21\u578b\u7d04 23 GB \u78c1\u789f\u30012.9 GB \u6d3b\u52d5\u8a18\u61b6\u9ad4\uff1b8B \u6a21\u578b\u7d04 4.2 GB \u78c1\u789f\u30011.0 GB \u6d3b\u52d5\u8a18\u61b6\u9ad4\u3002<\/li>\n\n\n\n<li><strong>\u786c\u9ad4\u9650\u5236<\/strong>\uff1a\u76ee\u524d\u50c5\u652f\u63f4 macOS Apple Silicon\uff0cCUDA \u4ecd\u5728\u898f\u5283\u4e2d\u3002<\/li>\n\n\n\n<li><strong>\u76ee\u6a19\u7528\u6236<\/strong>\uff1a\u60f3\u5728 M \u7cfb\u5217 MacBook \u4e0a\u672c\u6a5f\u8a66\u8dd1\u5927\u578b\u7a00\u758f\u6a21\u578b\u7684\u7814\u7a76\u8005\u3001\u958b\u767c\u8005\uff0c\u4ee5\u53ca\u5c0d VRAM \u4e0d\u5920\u3001\u53c8\u4e0d\u9858\u5b8c\u5168\u4f9d\u8cf4\u96f2\u7aef\u63a8\u8ad6\u7684\u5718\u968a\u3002<\/li>\n\n\n\n<li><strong>\u5dee\u7570\u5316<\/strong>\uff1a\u76f8\u8f03\u7d14\u91cf\u5316\u6216\u7d14\u5378\u8f09\u65b9\u6848\uff0cEdge0 \u628a prerouter \u9810\u6e2c\u8207 LoRA \u9084\u539f\u5305\u9032\u540c\u4e00\u689d\u7ba1\u7dda\uff0c\u8b93\u5378\u8f09\u5f8c\u7684\u7cbe\u5ea6\u640d\u5931\u6709\u660e\u78ba\u88dc\u511f\u8def\u5f91\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Edge0 \u5c6c\u65bc\u300c\u5148\u628a\u672c\u5730\u5927\u578b MoE \u8dd1\u8d77\u4f86\u300d\u9019\u500b\u52d9\u5be6\u8def\u7dda\u4e0a\u7684\u5de5\u5177\uff0c\u9650\u5236\u4e5f\u5f88\u6e05\u695a\uff1a\u5fc5\u9808\u5728 Apple Silicon \u4e0a\u8dd1\u3001\u9577\u4e0a\u4e0b\u6587 KV \u6703\u984d\u5916\u5403\u8a18\u61b6\u9ad4\u3001\u76ee\u524d\u6c92\u6709\u975e Apple \u5e73\u53f0\u7684\u5f8c\u7aef\u53ef\u7528\u3002\u5c0d\u60f3\u9a57\u8b49\u7a00\u758f\u6a21\u578b\u5728\u6d88\u8cbb\u786c\u9ad4\u53ef\u884c\u6027\u7684\u4eba\uff0c\u9019\u5957\u6846\u67b6\u7701\u53bb\u4e86\u81ea\u884c\u6574\u5408\u5378\u8f09\u8207\u9069\u914d\u5668\u9084\u539f\u7684\u529f\u592b\uff1b\u5c0d NVIDIA \u7528\u6236\uff0c\u5247\u8981\u7b49 CUDA \u5f8c\u7aef\u843d\u5730\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/Edge0-AI\/Edge0\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Edge0 \u662f\u4e00\u5957\u958b\u6e90\u4e32\u6d41 MoE \u63a8\u8ad6\u6846\u67b6\uff0c\u9760 SSD \u5c08\u5bb6\u5378\u8f09\u642d\u914d LoRA \u9084\u539f\u8207\u524d\u7f6e\u8def\u7531\u5668\uff0c\u5728 Apple Silicon \u4e0a\u628a 35B \u7b49\u7d1a\u7684 MoE \u6a21\u578b\u58d3\u5230 2.9 GB \u6d3b\u52d5\u8a18\u61b6\u9ad4\u5c31\u80fd\u8dd1\uff0c\u5c0d\u60f3\u5728\u7b46\u96fb\u672c\u6a5f\u73a9\u5927\u578b\u7a00\u758f\u6a21\u578b\u7684\u4eba\u4f86\u8aaa\u662f\u500b\u52d9\u5be6\u7684\u9078\u64c7\u3002<\/p>\n","protected":false},"author":8,"featured_media":12124,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,76,185,179,197,38,166,187],"tags":[],"class_list":["post-12125","post","type-post","status-publish","format-standard","hentry","category-133","category-76","category-qwen","category-nvidia","category-framework","category-38","category-mac","category-187"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12125","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=12125"}],"version-history":[{"count":2,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12125\/revisions"}],"predecessor-version":[{"id":12129,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12125\/revisions\/12129"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/12124"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=12125"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=12125"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=12125"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}