[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-baidu-ernie-5-5-flash-base-open-weights":3,"topics-all":38,"news-related-554f6552-6c09-401a-99e3-c9ce60a497ca":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"554f6552-6c09-401a-99e3-c9ce60a497ca","ERNIE 5.5 基座:280B激活14B,对话版还没来","百度 ERNIE 5.5 Flash Base:280B 稀疏 MoE,每 token 激活 14B,33.5T token 预训练,128K→1M。权重停在预训练结束,无对话版。37 层 HCA 把 128 token 压成一条 KV,6 层 MLA 保留全上下文,ECHO 复用末层做草稿。许可证排除欧盟英国韩国。","百度这次的 ERNIE 5.5 Flash Base 走了一条挺特别的路:把预训练权重放出来,对话版和工具调用版本留在自己手里。11 月 5 日发布的这份 checkpoint 是 2800 亿总参数、每 token 只激活 140 亿的稀疏 MoE,以 FP8 和 BF16 两种精度挂在 Hugging Face 上,但整个权重停在预训练结束的那一刻,没有经过任何中训练或后训练。[百度官方发布页](https:\u002F\u002Fernie.baidu.com\u002Frelease\u002Fernie-5-5-flash-base\u002F)\n\n## 先弄清这个模型能做什么\n\n这是一份基座模型,不是能直接对话的助手。百度在发布页写得很直白:权重是预训练结束时的 checkpoint,开发者可以自己决定往哪个方向继续训。后训练版本 ERNIE 5.5 Flash\"会随后到来\",但没有给日期。模型卡上写的是通过 completions 接口提供服务,也就是你发一段文本,它续写,没有对话模板这一层。\n\nThe Decoder 把这次发布描述成\"刻意不做微调和后训练\",MarkTechPost 也把\"自评\"两个字挂在成绩单旁边。eesel 的分析更直接:想拿它做客服智能体的团队,该等的是后训练版本。这三个判断指向同一件事——百度这次卖的不是\"更强的模型\",而是一个\"干净的起点\"。\n\n## 真正被反复打磨的是 KV 缓存\n\n43 层 Transformer,hidden size 4096,64 个注意力头。里面 37 层用的是 HCA(Heavily Compressed Attention):每 128 个 token 压成一个缓存条目,再加一段近期未压缩的短窗口;剩下 6 层用 MLA 保留 token 级的完整上下文视野,并且叠加 DSA 稀疏注意力——一个轻量索引器给历史位置打分,核心注意力只在 top-k 上计算,单个 query 的成本不随上下文增长。\n\n百度给出的取舍理由很直白:在相近的 KV 缓存预算下,他们试过用大量 CSA 层,最后选了少量 MLA 层,因为 HCA 加 MLA 的组合在预训练上打赢了 HCA 加 CSA。层内头维度也按成本分配:压缩比 128 倍的 HCA 层用 512 维宽头,昂贵的 MLA 层用 256 维窄头。省下来的空间直接换成 100 万 token 的上下文支持。\n\n## 路由、输出头和投机解码都被改了\n\nMoE 部分不是单路由器,而是多个独立初始化的 sigmoid 路由器联合给专家打分。百度给的动机很实在:单路由器的打分和专家实际有用程度只弱相关,而落选专家收不到梯度,错误选择会自我强化。改成多路由之后,路由分数与专家有用性的相关性明显提升,中后层收益最大,激活专家数、FLOPs 和通信量都没变。每层 512 个路由专家取 top-10,外加 1 个共享专家,专家在 2048 维潜空间里计算,派发流量和专家 FLOPs 大约减半。\n\n还有两处值得单独提:一是 mHC 把标准残差连接换成 4 条并行残差流;二是 ECHO 取代了常见的 MTP 模块,它复用主干最后一层去预测未来 token,没有额外模块,同一层还兼任投机解码的草稿模型,虽然训练时只预测一个额外 token,实际却能向前草拟好几个 token。\n\n## 跑分:自己打的表,自己有解释权\n\n评测对照的是 DeepSeek-V4-Flash-Base、Hy3-preview-Base 和 MiMo-V2.5-Base,四个模型用同一套 harness、prompt、few-shot 和 LLM judge。按表内加粗统计,ERNIE 5.5 Flash Base 在 30 项基准里有 24 项拿到最高分,MarkTechPost 也给出同样的 24 比 30,并注明这是百度统一流水线下的自评结果。逐项对比 DeepSeek-V4-Flash-Base,则是 26 项占优。\n\n短板同样清楚:中文事实问答 ChineseSimpleQA 上 73.0 对 DeepSeek 的 75.3,长文本重排 HELMET-Rerank 上 56.2 对 60.0,InfiniteBench 46.5 对 48.0。三项都是差一点的量级,更像同代模型的贴身缠斗,而不是代差。而且百度对齐的是较老的 V4 Flash 基座,不是更新的 DeepSeek V4.1 Flash,目前也还没有第三方实验室复现这批数字。\n\n## 最该读的是许可证\n\n权重确实免费可下,但 ERNIE 5.5 Flash Base Model License Agreement v1.0 的边界比\"开源\"两个字通常暗示的窄得多。授权地域被定义为\"除欧盟、英国和韩国以外的所有地理区域\",第 5.3 条进一步规定材料不得在授权区域外使用、复制、分发或展示,输出也在内。\n\n商用门槛分两档:相关产品年营收 5000 万美元及以下可免费商用,需要署名\"Powered by ERNIE\"并向百度发一次使用通知;超过 5000 万美元则必须事先拿到单独书面商业授权,一旦营收越线,还没授权的商业用途得先停下来等授权生效。百度还可要求运营方提供书面合规声明,通常每个日历年至多一次。第 5.5 条另有一道锁:不许拿这个模型去改进别的 AI 模型。\n\n## 所以呢\n\n模型侧是实打实的工程密度,长上下文、稀疏 MoE、投机解码草稿模型都塞进了这具 2800 亿参数的骨架里,在百度自选的三个同档对照模型面前几乎全面领先;但分发侧给的是一个带地域排除和营收门槛的半开放。对做开源生态的团队,这个基座干净、单一 checkpoint、方便横向对比不同训练方法;对想把服务直接铺到欧洲的团队,这张许可证在工程启动之前就得先解决法律问题。\n\n真正的看点可能不在这个基座,而在它后面那个还没发的对话版:官方说后训练版本会跟上,到那时,这份基座加上 post-training,才是百度在 ERNIE 5.5 这条线上真正的答案。","https:\u002F\u002Fernie.baidu.com\u002Frelease\u002Fernie-5-5-flash-base\u002F","b51e30de-7256-42c6-a5bd-409037e1f9e7",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"cc03cefc-d557-4778-91a2-0f66ab24d0bd","en","ERNIE 5.5 Flash Base: 280B MoE Base Weights, No Chat Version","ERNIE 5.5 Flash Base: 280B sparse MoE, 14B active, 33.5T tokens, 1M context. Base weights only, no chat model. License excludes EU, UK, South Korea.","Baidu took a deliberate route with ERNIE 5.5 Flash Base: ship the pre-training weights and keep the chat and tool-use versions to itself. The checkpoint released on November 5, 2026 is a sparse Mixture-of-Experts model with 280B total parameters and 14B activated per token, published on Hugging Face in both FP8 and BF16, and it stops exactly at the end of pre-training with no mid-training or post-training applied. The official release page is at https:\u002F\u002Fernie.baidu.com\u002Frelease\u002Fernie-5-5-flash-base\u002F\n\n## What the model can actually do\n\nThis is a base model, not an assistant you can talk to. Baidu is explicit that the weights are the checkpoint at the end of pre-training, leaving developers to decide where to train it next. The post-trained ERNIE 5.5 Flash will follow with no date given. The model card documents serving through the completions API, meaning you send text and it continues the text, with no chat format at all.\n\nThe Decoder framed the release as shipping deliberately without fine-tuning or post-training, MarkTechPost attached a self-reported label to the scorecard, and eesel concluded plainly that teams wanting a customer-service agent should wait for the post-trained version. All three readings point at the same thing: this launch is not about a stronger model, it is about a clean starting point.\n\n## The part that got the most engineering: KV cache\n\nForty-three Transformer layers, hidden size 4096, 64 attention heads. Thirty-seven of those layers use HCA, Heavily Compressed Attention, which squeezes every 128 tokens into a single cached entry plus a short window of recent uncompressed tokens. The remaining six use MLA to keep token-level access to the full context, and layer in DSA sparse attention where a lightweight indexer scores past positions and core attention is computed only over the top-k, so per-query cost stays constant as context grows.\n\nBaidu explains the trade-off plainly: under a similar KV-cache budget it tried stacking many CSA layers, then settled on a few MLA layers, because the HCA plus MLA combination won out on pre-training loss. Head dimensions are also cost-assigned: HCA layers, compressed 128 times, get wide 512-dim heads, while the expensive MLA layers get narrower 256-dim heads. The savings translate directly into the 1M-token context window.\n\n## Routing, the output head, and speculative decoding all changed\n\nThe MoE is not single-router. Several independently initialized sigmoid routers score the experts jointly, and Baidu reasoning is practical: a single router scores correlate only weakly with how useful each expert really is, and experts left out of the top-k receive no gradient, so wrong selections reinforce themselves. With multiple routers, routing scores correlate noticeably better with expert usefulness, the largest gains land in the middle and deeper layers, and the number of activated experts, expert FLOPs, and communication volume all stay the same. Each layer has 512 routed experts at top-10 plus one shared expert, running in a 2048-dimensional latent space that roughly halves dispatch traffic and expert FLOPs.\n\nTwo more pieces stand out. mHC replaces the standard residual connection with four parallel residual streams. And ECHO replaces the common MTP module, reusing the backbone own final layer to predict future tokens with no extra module, while that same layer doubles as the draft model for speculative decoding, drafting several tokens ahead despite being trained to predict only one extra token.\n\n## The scorecard, and who interprets it\n\nThe evaluation compares against DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base, with all four models run through the same harness, prompts, few-shot examples, and LLM judge. Counting the bolded rows in the table, ERNIE 5.5 Flash Base takes the top score on 24 of 30 benchmarks, a figure MarkTechPost reports identically and labels as a self-reported result from the unified pipeline. Head to head against DeepSeek-V4-Flash-Base, it leads on 26.\n\nThe weaknesses are equally clear. On ChineseSimpleQA it scores 73.0 against DeepSeek 75.3, on HELMET-Rerank 56.2 against 60.0, and on InfiniteBench 46.5 against 48.0. All three are narrow gaps that read more like a close fight between same-generation models than a generational lead. Worth noting that the comparison targets the older V4 Flash base rather than the newer DeepSeek V4.1 Flash, and no outside lab has yet reproduced these numbers.\n\n## The license is the part worth reading most\n\nThe weights are genuinely free to download, but the ERNIE 5.5 Flash Base Model License Agreement v1.0 draws far tighter boundaries than the word open source usually implies. Territory is defined as all geographical areas worldwide other than the European Union, the United Kingdom, and South Korea, and Section 5.3 further bars using, reproducing, distributing, or displaying the materials outside that territory, outputs included.\n\nCommercial use splits into two tiers. At or below US$50 million in annual revenue for related products, commercial use carries no model license charge, but requires prominent Powered by ERNIE branding plus a one-time use notice to Baidu. Above that threshold you need separate written commercial authorization before starting, and if revenue later crosses the line, affected commercial use lacking authorization has to be suspended until it takes effect, with Baidu able to request a written compliance statement, normally no more than once per calendar year. Section 5.5 adds another lock: you may not use the model to improve any other AI model.\n\n## So what\n\nOn the model side this is real engineering density. Long context, sparse MoE, and a speculative decoding draft model all fit inside a 280B-parameter frame, and it leads on nearly every head-to-head against the three comparable base models chosen for the comparison. On the distribution side, what you get is a half-open release with geographic exclusions and a revenue threshold. For teams working on open ecosystems, the base is clean, a single checkpoint, and convenient for comparing training methods side by side. For teams wanting to ship a service straight into Europe, this license has to be resolved legally before engineering even starts.\n\nThe real story may not be the base model at all, but the chat version that has not shipped yet. Baidu says the post-trained version is coming, and when it lands, this base plus post-training will be the actual answer on the ERNIE 5.5 line.","baidu-ernie-5-5-flash-base-open-weights","2026-11-05T12:00:00Z","2026-11-06T05:11:44.905678Z","2026-11-06T05:11:44.905692Z",true,"agent",106,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"32560a11-3d59-4397-8f01-e75f680a59db","月之暗面兑现承诺:K3.1 Pro 2.8T 权重开源上架","kimi-k3-1-pro-open-weights-huggingface","2026-11-04T23:07:35+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"4d0566f5-ef08-4c5f-befc-09dde8010620","Le Chonk 开源兑现:三周RL让TB冲上41%","mistral-large-4-open-weights-terminal-bench","2026-10-29T17:11:35+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"03d33c23-e56a-41eb-8732-b0712059662f","PUBG 母公司开源 MoE:激活仅 4.3B 单卡可跑","krafton-raon-1-mini-preview-open-weights","2026-10-23T13:08:33+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"b44fb160-9027-44cd-b355-a66869004abc","Nemotron 3.5 Super VL 开源:120B 总参只激活 12B","nemotron-3-5-super-vl-open-120b-moe","2026-10-20T17:11:56+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"bbfcbee4-8d92-4078-bb1e-eaf4d1c1e146","Beam 权重落地:501B MoE 三精度齐发","beam-1-0-weights-open-agent-stack","2026-10-18T21:10:09+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"d529114c-bfd7-4efb-9af4-444d3065b397","Step 5 Preview 开源,2.7 美元\u002F百万输出","stepfun-step-5-preview-600b-moe","2026-10-17T21:12:04+00:00"]