[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-cohere-command-a-218b-moe-2h100-63pct":3,"topics-all":36,"news-related-6def00c6-c2b8-46ba-b423-cc9b84ede534":55},{"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},"6def00c6-c2b8-46ba-b423-cc9b84ede534","Cohere Command A+：首个开源218B MoE大模型，两块H100即可部署","5月20日，Cohere发布Command A+，这是其首款完全开源权重的前沿级模型：2180亿总参数，稀疏MoE架构每个Token仅激活250亿参数，Apache 2.0许可、开源在Hugging Face上，提供FP8和W4A4量化版本，推荐配置仅需两块H100 GPU即可运行。相比上一代Command A Reasoning，吞吐量提升约63%——W4A4量化（专家参数4-bit、注意力路径保持高精度的近无损压缩）下可达每秒约375个输出Token，首Token延迟113ms，并支持MoE speculative decoding带来额外1.5-1.6倍加速。\n\n从benchmark看，Command A+在多个任务上实现了质的飞跃。代理任务τ²-Bench从37%跃升至85%，代理编程Terminal-Bench Hard从3%升至25%，MMMU多模态推理达75.1%、MathVista达80.6%。值得注意的是，该模型截止训练数据为2024年6月，并未依赖更大的前沿-scale数据，却实现了显著超越，这说明架构与训练效率的优化空间依然巨大。\n\n开源权重、可私有部署、W4A4量化让200B+模型首次在企业级硬件上可用，这对金融、医疗等有数据主权要求的行业意义重大。Command A+代表了2026年开源大模型的重要一步：不再追求参数堆砌，而是通过稀疏MoE架构和极致量化，让强大模型真正触手可及。MoE+量化的组合拳正在改写大模型的成本结构，小团队跑SOTA模型的时代正在加速到来。","https:\u002F\u002Fmer.vin\u002F2026\u002F05\u002Fcohere-command-a-open-source-218b-moe-llm-on-two-h100-gpus\u002F","df9f8204-8e8d-4fce-8526-3c6fe8e6ae56",[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},"33a79139-9c50-48f1-a671-c45a4faaee4c","en","Cohere Command A+: open 218B MoE, runs on two H100s","On May 20, Cohere released Command A+ — its first fully open-source-weight frontier-tier model: 218B total parameters, a sparse MoE architecture activating only 25B per token, Apache 2.0 licensed, open-sourced on Hugging Face, with FP8 and W4A4 quantization versions. The recommended setup is just two H100 GPUs. Compared to the previous-generation Command A Reasoning, throughput improves by roughly 63% — under W4A4 quantization (4-bit for expert parameters, near-lossless compression with the attention path kept at high precision), it reaches about 375 output tokens per second, with 113ms first-token latency. MoE speculative decoding adds another 1.5-1.6× speedup.\n\nOn benchmarks, Command A+ achieves qualitative leaps across multiple tasks: the agent task τ²-Bench jumps from 37% to 85%; agent coding Terminal-Bench Hard climbs from 3% to 25%; multimodal reasoning MMMU hits 75.1% and MathVista 80.6%. Notably, the model's training data cutoff is June 2024 — it does not rely on larger frontier-scale data, yet achieves significant gains, indicating that the optimization space in architecture and training efficiency remains vast.\n\nOpen-source weights, private deployment, and W4A4 quantization make a 200B+ model available on enterprise-grade hardware for the first time — significant for industries with data-sovereignty requirements like finance and healthcare. Command A+ represents an important step for open-source large models in 2026: no longer chasing parameter stacking, but using sparse MoE architecture and extreme quantization to make powerful models truly accessible. The MoE + quantization combo is rewriting the cost structure of large models; the era of small teams running SOTA models is arriving faster than expected.","cohere-command-a-218b-moe-2h100-63pct","2026-05-22T11:06:00Z","2026-05-22T19:07:05.553961Z","2026-08-19T02:08:40.142862Z",true,"agent",425,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"559df680-f13e-48f0-9263-73116aaff411","上交大DLite免历史投机解码:Agent提速2.74倍","dlite-history-free-speculative-decoding","2026-10-19T17:07:52+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"2638aeac-dc4d-4b73-b7fe-2b042015adee","OreoLook 开源:三层缓存把 AI 搜索搬进 8 核 CPU,重复问题 0.1 毫秒出答案","oreolook-three-layer-cpu-cache","2026-09-10T23:08:36+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"0fe9ceb8-6411-4924-8e02-8cee3665fc6f","Cohere 开源 megakernel 推理引擎：单 CUDA 文件，H100 解码吃到 62% 带宽光速","cohere-megakernel-north-mini-code-h100","2026-09-08T21:13:46+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"0c29e1ad-914a-4b79-a153-445c087acb03","被 LLM 抛弃的 dropout 翻身:Cerebras 称调好可省 25% 训练 FLOPs","dont-drop-dropout-layer-sparsity","2026-09-07T21:06:35+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"199cd4ef-f092-45a5-8635-91778dd2bce2","编译即训练：一句规约炼出 83.6% 准确率的本地神经函数，教师模型只用一次","compile-by-training-neural-functions","2026-09-04T23:08:03+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"7a6d28b6-65da-4a29-96b1-dedb9894de97","随机驱逐追平最强打分器:Salesforce 重写 KV Cache 压缩常识","random-attention-kv-cache-eviction","2026-09-04T19:08:26+00:00"]