[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-vidraft-vkue-cpu-moe":3,"news-related-518d0be2-1220-4920-98fd-dee9df27a43d":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"518d0be2-1220-4920-98fd-dee9df27a43d","VIDRAFT VKUE 把同一份 34.7B 稀疏 MoE 权重从 B200 跑到裸 CPU","来自 FINAL-Bench 的 VIDRAFT 团队本周在 Hugging Face 社区博客发布 VKUE(VIDRAFT Kernel Ubiquitous Engine)实测,把同一份 Ourbox-35B-JGOS 权重从单卡 B200 数据中心(聚合 18,057 tok\u002Fs)、单卡 A10G(126 tok\u002Fs)、8GB 显存笔记本(20 tok\u002Fs),一路压到完全无 GPU 的 CPU 服务器(~17 tok\u002Fs)。每个数字都带公开复现路径和 GPU\u002FCPU 双路在线 demo,不是 PPT 上的承诺。\n\n关键洞察不在新 kernel,而在模型本身的物理性质:Ourbox-35B-JGOS 来自 Qwen3.5-MoE \u002F Qwen3-Next 系列,34.7B 总参数但每次 token 仅激活约 3B(256 专家 top-8,Gated-DeltaNet 线性注意力与全注意力交错)。解码是 memory-bandwidth bound,单 token 实际搬动约 1.45 GB,比同体积密集 34B 的 16.7 GB 缩小 11 倍——这才是「同一份权重能塞进 8GB 卡」的根本原因。\n\n在同一台 8GB 笔记本、同一 VKUE 引擎、同一 Q3_K_M 量化下做严格 A\u002FB:稀疏 A3B 拿到 20.01 tok\u002Fs,Qwen2.5-32B 密集基线只跑出 5.36 tok\u002Fs,3.7× 加速完全来自稀疏本身。能力侧 Ourbox-35B 在 GPQA Diamond 上拿到 86.4%(maj@8)\u002F 70.7%(greedy)。团队口号「VKAE 求快,VKUE 求广」点出工程转向——sparse MoE 的红利不该被数据中心独享,自托管、边缘部署和公共部门场景都能用上 frontier-class reasoner。","https:\u002F\u002Fhuggingface.co\u002Fblog\u002FFINAL-Bench\u002Fvkue","24d5c6c5-6573-4180-a1fd-f1459842d1af",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"37d9b12c-cd2f-41b1-838f-de3b282e017c","en","VKUE runs the same 34.7B sparse MoE from B200 to bare CPU","The VIDRAFT team from FINAL-Bench this week published VKUE (VIDRAFT Kernel Ubiquitous Engine) benchmark results on the Hugging Face community blog, running the same Ourbox-35B-JGOS weights across single-card B200 data center (aggregate 18,057 tok\u002Fs), single-card A10G (126 tok\u002Fs), 8GB-VRAM laptop (20 tok\u002Fs), all the way down to a completely GPU-less CPU server (~17 tok\u002Fs). Every number comes with a public reproduction path and live GPU\u002FCPU demo — not just PPT promises. The key insight isn't a new kernel, but the physical nature of the model itself: Ourbox-35B-JGOS comes from the Qwen3.5-MoE \u002F Qwen3-Next family, with 34.7B total parameters but only ~3B activated per token (256 experts top-8, Gated-DeltaNet linear attention interleaved with full attention). Decoding is memory-bandwidth bound: each token actually moves about 1.45 GB, 11× less than a same-sized dense 34B's 16.7 GB — that's the fundamental reason \"the same weights can fit in an 8GB card\". A strict A\u002FB on the same 8GB laptop, the same VKUE engine, and the same Q3_K_M quantization: sparse A3B gets 20.01 tok\u002Fs, the Qwen2.5-32B dense baseline only 5.36 tok\u002Fs — the 3.7× speedup comes entirely from the sparsity. On capability, Ourbox-35B gets 86.4% (maj@8) \u002F 70.7% (greedy) on GPQA Diamond. The team's slogan \"VKAE chases speed, VKUE chases breadth\" signals the engineering pivot — the dividends of sparse MoE shouldn't be monopolized by data centers, and self-hosting, edge deployment, and public-sector scenarios should all be able to use frontier-class reasoners.","vidraft-vkue-cpu-moe","2026-07-12T10:00:00Z","2026-07-16T06:15:22.622018Z","2026-08-19T02:08:40.142862Z",true,"agent",117,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"cac485ab-a429-4ebc-88c6-a1f924f978ff","AWS 开源 KeysAndValues:微调时就让模型学会“遗忘”,单张 A100 撑住 128K","aws-keysvalues-sparse-attention-finetuning","2026-08-26T05:20:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"c94bdf86-5de9-49fe-8c98-0f5c47611bfe","SGLang v0.5.18 发布:大模型冷启动提速 2.38 倍,710 个 PR 都改了什么","sglang-v0-5-18-cold-start-2-38x","2026-08-24T23:15:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"92433e6b-113a-4ada-af77-fbb8995a9850","LFM2.5-DSpark 开源:300M 草稿模型让端侧推理快 2.87 倍,输出零损耗","lfm2-5-dspark-draft-models","2026-08-21T21:10:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"926f89fc-5ed5-4170-bfac-931d3a31b6a4","腾讯混元开源 AngelSpec 投机解码框架：DFly 在 Hy3-A21B 上取得 1.98–2.40× 加速","tencent-angelspec-spec-decoding-hy3-dfly","2026-07-30T00:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"894787b7-8799-4c22-8e20-5e04db8f7181","DeepSeek V4 GA 把分时电价塞进大模型 API:CSA + HCA 撑起 1M 上下文,峰值翻倍只是商业答案","deepseek-v4-ga-time-of-use-pricing","2026-07-24T02:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"f8a33525-3846-46e2-9f88-05be407dbe2f","Liquid AI 用 Antidoom 把推理模型的 doom loop 压到 1%：一次训练只改 1 个 token","liquid-ai-antidoom","2026-07-12T10:00:00+00:00"]