[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-liquid-ai-lfm3-on-device-family":3,"topics-all":38,"news-related-4d5e175c-8472-446e-acdb-e9f1c6504b6c":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},"4d5e175c-8472-446e-acdb-e9f1c6504b6c","Liquid AI 开源 LFM3:六模型凑齐端侧全模态","10 月 27 日 Liquid AI 发布 LFM3:六个开放权重 2B 级模型共用同一个 1.7B 底座,覆盖文本、视觉、音频、全模态与两个决策模型。文本版 22T token 预训练、256K 上下文,4-bit 量化后约 1GB,官方基准中 agent 平均分压过 Qwen3.5-9B。","端侧模型这条赛道,大多数玩家的思路是\"把一个大模型裁小塞进手机\"。Liquid AI 在 10 月 27 日给出了另一个答案:LFM3 系列——六个 2B 级开放权重模型,全部从同一个 LFM3-1.7B-Base 底座分叉而来,覆盖文本、视觉、音频、全模态(omni)和两个决策模型,权重当天上架 Hugging Face。\n\n## 一个底座,六个模型\n\nLFM3-1.7B-Base 用 22T token 完成预训练,最后 3.5T token 是精选数据的中期训练,把上下文窗口撑到 256K。家族成员各管一摊:文本模型 LFM3-1.7B 面向本地智能体;视觉模型 LFM3-VL-2B 是 LFM 系列视觉模型首次支持视频输入,还会逐步推理再作答;音频模型 LFM3-Audio-2B 支持 18 种语言的语音对话与工具调用;LFM3-Omni-2B(2.2B)是公司首个全模态模型,能看、能听、能说。另有两个决策模型 d1-2B 和 d1-Omni-2B:不写长答案,一次前向传播直接给出带概率标定的判断,d1-2B 还比上代 d1-3B 少三分之一参数、基准分反升。\n\n## 三混合架构,一半以上是卷积\n\n底座混合了三种算子:24 层里 13 层是门控短卷积(kernel size 从 3 提到 7),5 层是 256-token 窗口的滑动注意力,6 层是全局注意力。这次架构搜索的硬件在环范围从上代的一台手机加一台笔记本,扩到 14 种设备与后端组合,Galaxy S26 Ultra、树莓派 5、Apple M4 Max、AMD Ryzen AI、高通 Hexagon NPU 都在列。\n\n## 数字怎么说\n\n官方基准表里最能打的是智能体能力:LiveClaw 45.60,上一代 LFM2.5-2.6B 只有 14.70,Qwen3.5-9B 是 35.77;全表平均 49.37,同样压过 Qwen3.5-9B 的 47.24——参数约为对方的六分之一。视觉侧,LFM3-VL-2B 在 62 项视觉基准上平均 71.4 分,与体积六倍于它的 Gemma 4 12B(71.7)基本打平。端侧数据:4-bit 量化后约 1GB,llama.cpp 实测三台设备解码速度是次名模型的 1.8-2.8 倍,128K 上下文的 KV cache 只要 1.6GB(Qwen3.5-4B 要 4.3GB)。云端一张 B300 跑 SGLang 每天可产出 36 亿 token,官方甚至用 30 万个并行智能体跑了一遍 Enron 邮件库。\n\n## 两个保留意见\n\n其一,以上分数基本来自官方发布材料,独立复测还需要时间,目前外部报道也主要引用发布稿。其二,官方自己承认在智能体编程和知识问答上大模型仍占优——GPQA Diamond 63.37 对 Qwen3.5-9B 的 80.40,差距明显。所以 LFM3 的赌注不是\"小模型全面胜出\",而是专用化路线:2B 级足够小,单卡 A100 几小时微调就能造就一个垂直领域专家,常驻设备端运行。MacPaw 的实测是个佐证:65 个 macOS 真实智能体任务上,它对 Qwen3.5-2B 打出 64% 对 48% 的成功率,token 还省 58%。\n\n当云端玩家还在卷万亿参数,Liquid AI 把战场挪到了\"每台设备里住几个小专家\"。留给读者的问题是:你的下一个应用,需要一位通才,还是六个专才?\n\n参考:[LFM3 官方发布](https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm3) | [AI Weekly 报道](https:\u002F\u002Faiweekly.co\u002Falerts\u002Fliquid-ai-opens-six-model-lfm3-family-at-256k-context)","https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm3","511bb1e6-a31f-4dc1-929b-9a7582e67447",[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},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"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},"bb587418-b3c8-4c55-a273-bb08e89fe3a9","en","Liquid AI Open-Sources LFM3: Six Models, One Backbone","On Oct 27 Liquid AI released LFM3: six open-weight 2B-class models sharing one 1.7B backbone, spanning text, vision, audio, omni and decision tasks.","The edge-model race usually means shrinking one big model to fit a phone. On October 27, Liquid AI offered a different answer: LFM3, a family of six open-weight models in the 2B class, all branched from a single LFM3-1.7B-Base backbone, covering text, vision, audio, omni, and two decision models. Weights went live on Hugging Face the same day.\n\n## One backbone, six models\n\nLFM3-1.7B-Base was pre-trained on 22T tokens; the final 3.5T tokens form a mid-training stage on curated data that extends the context window to 256K. Each family member has a job: LFM3-1.7B (text) targets local agents; LFM3-VL-2B (vision) is the first LFM vision model to take video input, and it reasons step by step before answering; LFM3-Audio-2B supports voice chat and tool calling across 18 languages; LFM3-Omni-2B (2.2B) is the company's first omni model that can see, listen, and talk back. Two decision models, d1-2B and d1-Omni-2B, skip long answers and return calibrated judgments in a single forward pass; d1-2B replaces d1-3B with a third fewer parameters while scoring higher.\n\n## A tri-hybrid architecture, over half convolution\n\nThe backbone mixes three operators: of 24 layers, 13 are gated short convolutions (kernel size raised from 3 to 7), 5 are sliding-window attention with a 256-token window, and 6 are global attention. Hardware-in-the-loop architecture search expanded from one phone and one laptop for LFM2 to 14 device and backend combinations, including the Galaxy S26 Ultra, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI, and the Qualcomm Hexagon NPU.\n\n## What the numbers say\n\nThe agent benchmarks are the strongest part: LFM3-1.7B posts 45.60 on LiveClaw versus 14.70 for LFM2.5-2.6B and 35.77 for Qwen3.5-9B; its table average of 49.37 also beats Qwen3.5-9B's 47.24, with roughly one-sixth the parameters. On the vision side, LFM3-VL-2B averages 71.4 across 62 vision benchmarks, essentially matching Gemma 4 12B (71.7), a model more than six times larger. On-device: roughly 1 GB at 4-bit, decoding 1.8-2.8x faster than the next model across three devices under llama.cpp, and a 128K-token KV cache of just 1.6 GB (versus 4.3 GB for Qwen3.5-4B). On a single NVIDIA B300 with SGLang it can produce 3.6 billion tokens per day; the team even ran more than 300,000 parallel agents through the Enron email archive.\n\n## Two caveats\n\nFirst, these scores come from the official launch post; independent replication will take time, and outside coverage so far mostly cites the announcement. Second, Liquid AI itself admits larger models still win on agentic coding and knowledge: GPQA Diamond 63.37 versus 80.40 for Qwen3.5-9B is a clear gap. So the LFM3 bet is not \"small models win across the board\" but specialized intelligence: a 2B-class model is small enough to fine-tune into a domain expert on a single A100 in hours, then run on-device indefinitely. MacPaw's test supports the thesis: across 65 real macOS agent tasks, it beat Qwen3.5-2B with 64% versus 48% task success while using 58% fewer tokens.\n\nWhile cloud players chase trillion-parameter scale, Liquid AI moved the battlefield to a few small experts living in every device. The question for builders: does your next app need one generalist, or six specialists?\n\nReferences: [LFM3 launch post](https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm3) | [AI Weekly coverage](https:\u002F\u002Faiweekly.co\u002Falerts\u002Fliquid-ai-opens-six-model-lfm3-family-at-256k-context)","liquid-ai-lfm3-on-device-family","2026-10-27T23:20:00Z","2026-10-27T23:08:06.172454Z","2026-10-27T23:08:06.172475Z",true,"agent",229,[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},"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":65,"title":66,"news_slug":67,"published_at":68},"faf5c85e-ed83-432a-8b55-80c9356862c5","腾讯优图开源5B全模态解析:文档榜96.96超GPT-5.2","youtu-parsing-omni-5b","2026-10-19T19:09:27+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"ac666c4e-4756-43d8-840a-4eacbf88f804","Meta 开源 816B MoE 模型 Muse Spark Open,跑在两台 Mac Studio 上","meta-muse-spark-open-816b-mac-studio","2026-10-17T03:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"1b6ec126-cba1-47da-870b-da19cc523763","Meta 开源 816B MoE 模型 Muse Spark Open:1M 上下文,两台 Mac Studio 跑得动","meta-muse-spark-open-816b-moe-1m-mac-studio","2026-10-15T11:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"5d4acf64-67d8-404b-9d3f-cdf7761fb8ac","小鹏开源 GroundingPI:4B 视觉 grounding 模型自评超 GPT-6 Astra","groundingpi-4b-visual-grounding-foundation-model","2026-10-13T17:10:26+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"cd49f913-cde7-4cf3-8d93-24508653180e","腾讯混元开源AuK:1.5B语音模型统一生成与编辑,4步推理快4.5倍","tencent-hunyuan-auk-speech-editing","2026-09-09T09:12:00+00:00"]