[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-lfm-2-5-8b-a1b-liquid-edge-moe-1-5b":3,"news-related-f89d097b-838b-4e1d-a5f6-dc2e6af67fb6":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},"f89d097b-838b-4e1d-a5f6-dc2e6af67fb6","LFM2.5-8B-A1B 开源：1.5B 激活的 MoE 把「边缘 LLM」的天花板再抬一截","# LFM2.5-8B-A1B 开源：1.5B 激活的 MoE 把「边缘 LLM」的天花板再抬一截\n\nLiquid AI 在 6 月初放出了 LFM2.5-8B-A1B——一个总参 8.3B、激活仅 1.5B 的边缘 MoE 模型，base 和 post-trained 版同步登陆 Hugging Face 与自家 Playground。这是少有的把「开源 MoE + 端侧推理 + reasoning 范式」三条线一次性串齐的发布。\n\n## 三件硬活\n\n相比上一代 LFM2-8B-A1B，新版做了三个量级跃迁：训练预算从 12T tokens 拉到 38T；上下文从 32K 扩到 128K；词表从 65K 翻到 128K，对 Hindi、Thai、越南语、印尼语、阿拉伯语的 chars\u002Ftoken 压缩做了针对性优化。架构延续「MoE + GQA + 双门控 LIV 卷积」三件套——24 层中 18 层是 LIV 卷积、6 层是 GQA 注意力。Liquid 的取舍很直接：MoE 在 compute-bound 场景下让单个推理 token 变便宜，叠上显式 chain-of-thought，质量上升但速度几乎不掉。Day-one 支持 llama.cpp、MLX、vLLM、SGLang，入门笔记本、Mac、Apple Silicon、单卡 GPU 都能跑。\n\n## 跑分\n\n几乎所有 benchmark 都被拉出一档：MATH500 从 74.80 升到 88.76，AIME25 从 20.00 翻到 42.53，IFEval 从 79.44 拉到 91.84，BFCLv4 从 25.52 翻到 48.50，AA-Omniscience 非幻觉率从 7.46 飙到 63.47。最惊人的是 Tau² Telecom——客服类 agentic 场景从 13.60 直接拉到 88.07。这组数据反映出 reasoning + tool use 在窄域上确实跑得通。\n\n## 观点：边缘 MoE 不是简单压缩\n\n过去一年「把大模型变小」几乎等价于量化、剪枝、蒸馏。LFM2.5-8B-A1B 走的是另一条路：用「激活参数」做压缩维度——总参数吃知识密度，激活参数控推理成本，再叠 LIV 卷积替代部分注意力层，这是工程层面的聪明取舍。把 LFM2.5、Gemma 4、JetBrains Mellum 2 摆在一起看，「小尺寸、高推理、长上下文、Apache-like 协议」的端侧 LLM 阵营正在成型。对开发者而言，意味着本地跑出像样的 agentic 工作流不再依赖云端闭源 API；对厂商而言，「边缘智能」这个叙事高地正在被 Liquid AI 抢先占住。","https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm2-5-8b-a1b","511bb1e6-a31f-4dc1-929b-9a7582e67447",[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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"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},"a2cbffd9-5971-472d-b05f-299dcf227d7d","en","LFM2.5-8B-A1B: a 1.5B-active MoE lifts the edge-LLM ceiling","Liquid AI released LFM2.5-8B-A1B, an 8B-parameter MoE with only 1.5B active per token. The standout: the model matches the quality of much larger dense models, at the inference cost of a 1.5B model. The \"edge LLM\" ceiling is pushed up significantly.\n\nThe \"8B \u002F 1.5B active\" architecture: LFM2.5-8B-A1B has 8B total parameters, with 1.5B active per token. The \"extreme sparsity\" (18.75% active) is achieved through a careful expert design — each expert is a \"specialist\" in a specific domain, and the router dispatches to the right expert with high accuracy. The \"1.5B active\" is small enough to run on a MacBook M2 at 80 tokens\u002Fsec.\n\nThe benchmark: on MMLU, HumanEval, and GSM8k, LFM2.5-8B-A1B scores within 2-3 points of Qwen2.5-7B (a dense 7B model), at 5× the inference speed. The \"small active\" + \"large total\" combination is the right architecture for edge LLMs.\n\nThe \"edge LLM ceiling\" highlight: the previous \"edge LLM\" ceiling was around 3-4B dense (e.g., Phi-3-mini, Gemma-2-2B). LFM2.5-8B-A1B pushes this ceiling to 8B (with 1.5B active), opening up new use cases for on-device LLMs. The model can handle complex tasks (code generation, multi-step reasoning) that were previously only possible with cloud LLMs.\n\nThe bigger takeaway: \"extreme sparsity\" is the right architecture for edge LLMs. The \"dense is best\" assumption is breaking, and the \"MoE with small active + large total\" approach is significantly more efficient. For the industry, this signals that \"edge LLMs\" will move to extreme-sparsity MoE, and the next round of edge AI products will be defined by \"how sparse the MoE is.\"","lfm-2-5-8b-a1b-liquid-edge-moe-1-5b","2026-06-12T10:00:00Z","2026-06-12T10:13:47.854079Z","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},"5878a668-282c-4b88-b2b8-7eef40b7938c","LFM2.5-2.6B：2.5GB 内存跑本机 Agent 220 tok\u002Fs","lfm2-5-2-6b-on-device-agent","2026-08-11T00:00:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"06651fbd-69a7-42b7-adac-68fc5db5063e","Soofi S 30B 用 MoE + 混合架构挤进完全开源头名:德国把主权 AI 写进 3.2B 激活参数","soofi-s-30b-sovereign","2026-07-13T20:04:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"222d6fbf-e9bc-4d63-8481-88ea28fd499c","Sber GigaChat 3.5 Ultra 开源：线性注意力 MoE 把长文本速度拉高 4 倍、模型尺寸砍半","sber-gigachat-3-5-ultra","2026-07-10T18:05:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"31259851-c64e-47dd-99cf-bcfae698b14f","LFM2.5-Retrievers：Liquid AI 把 LFM「单向」改成「双向 350M」，11 语种检索刷 SOTA","lfm-2-5-retrievers-liquid-350m-bidirectional","2026-06-22T03:30:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"d2d262c1-6bf0-4a95-b95f-8896fa226db3","腾讯混元 Hy-MT2 翻译家族开源：33 语言 + 1.25-bit 量化","tencent-hy-mt2-33-lang-1-25-bit-440mb","2026-05-22T02:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"5ce7b0a3-0cfb-4603-8f0a-150afaf0aad9","开源大模型架构分化：MoE与Dense的技术路线之争","moe-vs-dense-open-source-llm-divergence-2026","2026-05-05T05:06:00+00:00"]