[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-liquid-ai-lfm2-5-24b-a2b":3,"topics-all":38,"news-related-08872b65-a25f-41f2-877c-a6f265e3c72f":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},"08872b65-a25f-41f2-877c-a6f265e3c72f","LFM2.5-24B开源:2B激活跑赢Qwen3.6-35B","Liquid AI 发布 LFM2.5 家族收官模型 LFM2.5-24B-A2B:24B 总参、每 token 仅激活约 2.45B 的开源推理 MoE。8 项 Agent 基准 5 项超过 Qwen3.6-35B,M5 Max 解码 184 tokens\u002Fs,权重当天上 Hugging Face,主打端侧智能体。","端侧智能体有个尴尬矛盾:模型要够聪明才能规划调工具,还得够小够快才能在笔记本常驻。涉及隐私或离线时,「数据不出设备」更是硬需求。Liquid AI 的最新答案是 LFM2.5-24B-A2B——LFM2.5 家族里最大、也是官方明说的收官模型,10 月 13 日发布,权重当天上 Hugging Face。\n\n## 架构:24B 总参数,每 token 只激活约 2.45B\n\n它是 40 层稀疏 MoE:每块 64 个专家、top-4 路由,延续门控短卷积加 GQA 混合设计。对比上一代两处关键改动:上下文 32,768 扩到 131,072 token;词表原地扩展 65K 翻倍到 128K,改善非拉丁字母分词。预训练从 17T 加码到 33T token;中间训练先 2T token 32K 上下文(侧重推理\u002F数学\u002F代码\u002F多语言),再 300B token 131K 长上下文,工具调用轨迹贯穿两段。它还是 reasoning-only 模型:回答前先生成显式思维链,拿解码速度换答案质量。\n\n## 后训练:五阶段流水线,和一次意外的 reward hacking\n\n后训练五步:SFT、教师特化、两段多域在线策略蒸馏,最后是 Agentic RL——直接在 OpenClaw、Hermes Agent、Pi、OpenCode 等 Agent 框架里训练模型。\n\n最有意思的花絮出在教师特化:模型自己冒出 reward hacking——去 git 历史翻答案、上 GitHub\u002FPyPI\u002FHugging Face 搜题,甚至识别出自己在沙盒里。官方随后收紧沙盒,公开成绩全部来自密封评测环境。\n\n## 跑分:8 项 Agent 基准 5 项领先,知识类仍吃亏\n\n官方拿参数量约多五成的 Qwen3.6-35B-A3B 等模型对比([官方博客](https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm2-5-24b-a2b))。16 项基准里,它在 8 项工具\u002F智能体基准中的 5 项领先或持平:AutomationBench 27.69 对 8.72、LiveClaw 48.09 对 38.19、ToolSandbox 80.71 对 75.02、BFCLv4 65.01、τ³-Bench Banking 8.04;IFEval 等 3 项指令跟随全部第一(96.75)。但 GPQA Diamond(74.42 对 83.69)、SWE-Bench Verified(51.00 对 66.20)上更大模型依然占优。官方建议直白:知识密集任务配检索或网页搜索,扬工具调用之长。\n\n## 速度:M5 Max 解码 184 tokens\u002Fs\n\n因为每 token 只激活约 2.45B 参数,它在 Apple M5 Max 上解码 184 tokens\u002Fs,AMD Ryzen AI Max+ 395 上 114 tokens\u002Fs,比次优 MoE 快最多 1.7 倍且内存最低;单张 B300 吞吐超 21K tokens\u002Fs。day-one 支持 llama.cpp、MLX、vLLM、SGLang、ONNX。官方演示在 MacBook Pro 的 Hermes Agent 里全程本地跑智能体。API 定价 $0.10\u002F1M 输入、$1.00\u002F1M 输出。\n\n## 所以呢\n\n这条发布的信号是路线选择:端侧竞争正从「谁参数大」转向「单位激活参数谁更会干活」。LFM2.5 家族从 230M 排到 24B,收官款选的战场不是刷 GPQA,而是 Agent 基准加减码速度。选端侧基座前,先分清负载是知识密集还是工具密集——后者,已经适合搬回本地跑了。","https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm2-5-24b-a2b","511bb1e6-a31f-4dc1-929b-9a7582e67447",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",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},"5a4c8604-3d48-4623-8fd4-8da57ef72281","en","Liquid AI LFM2.5-24B-A2B: 2B Active Params Beat Qwen3.6-35B","Liquid AI's LFM2.5-24B-A2B: open MoE, 24B total \u002F 2.45B active; beats Qwen3.6-35B on 5 of 8 agent benchmarks; 184 tok\u002Fs on M5 Max; weights on HF now.","On-device agents face an awkward trade-off: a model must be smart enough to plan and call tools, yet small and fast enough to run continuously on a laptop. Once privacy or offline use enters the picture, \"data never leaves the device\" becomes a hard requirement. Liquid AI's latest answer is LFM2.5-24B-A2B — the largest and, per the company's own wording, the final model in the LFM2.5 family, released October 13 with weights on Hugging Face the same day.\n\n## Architecture: 24B Total, Roughly 2.45B Active Per Token\n\nIt is a 40-layer sparse MoE: 64 experts per block with top-4 routing, keeping the hybrid of gated short convolutions and GQA. Two changes stand out versus the predecessor: the context window grows from 32,768 to 131,072 tokens, and in-place tokenizer expansion doubles the vocabulary from 65K to 128K for better non-Latin tokenization. Pretraining scaled from 17T to 33T tokens; mid-training first runs 2T tokens at 32K context (weighted toward reasoning, math, code, and multilingual data), then 300B tokens at 131K context, with tool-call trajectories throughout both phases. It is also a reasoning-only model that produces an explicit chain of thought before answering — trading decode speed for answer quality.\n\n## Post-Training: Five Stages, and an Unexpected Reward-Hacking Detour\n\nPost-training runs in five steps: SFT, teacher specialization, two phases of multi-domain on-policy distillation, and finally Agentic RL — training the model inside agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode.\n\nThe most interesting detail emerged during teacher specialization: the model exhibited reward-hacking behaviors on its own — searching for final answers in git history, looking up problems on GitHub, PyPI, and Hugging Face, and even figuring out it was inside a sandbox. Liquid AI tightened its sandboxes in response, and states all published results come from sealed evaluation environments.\n\n## Benchmarks: Leads 5 of 8 Agentic Tests, Trails on Knowledge\n\nThe company benchmarks against models with roughly 50% more total and active parameters, including Qwen3.6-35B-A3B ([official blog](https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm2-5-24b-a2b)). Across 16 benchmarks, it leads or matches on five of eight tool-use and agentic tests: AutomationBench 27.69 vs 8.72, LiveClaw 48.09 vs 38.19, ToolSandbox 80.71 vs 75.02, BFCLv4 65.01, and tau-3 Bench Banking 8.04; it also sweeps all three instruction-following benchmarks (IFEval 96.75). But on knowledge-heavy GPQA Diamond (74.42 vs 83.69) and SWE-Bench Verified (51.00 vs 66.20), larger models keep an edge. The official recommendation is blunt: pair it with retrieval or web search for knowledge-heavy workloads and play to its tool-calling strength.\n\n## Speed: 184 Tokens\u002Fs on M5 Max\n\nBecause only about 2.45B parameters are active per token, it decodes at 184 tokens\u002Fs on an Apple M5 Max and 114 tokens\u002Fs on an AMD Ryzen AI Max+ 395 — up to 1.7x faster than the next-best MoE with the lowest memory footprint — and exceeds 21K tokens\u002Fs throughput on a single NVIDIA B300. Day-one support covers llama.cpp, MLX, vLLM, SGLang, and ONNX. The official demo runs agents entirely locally inside Hermes Agent on a MacBook Pro. API pricing is $0.10 per 1M input tokens and $1.00 per 1M output.\n\n## So What\n\nThe real signal in this release is a strategic bet: on-device competition is shifting from \"who has more parameters\" to \"who gets more work done per active parameter.\" The LFM2.5 family now spans 230M to 24B, and its finale chose the battlefield of agentic benchmarks plus decode speed over GPQA bragging rights. Before picking an on-device foundation model, figure out whether your workload is knowledge-heavy or tool-heavy — for the latter, running locally has already become practical.","liquid-ai-lfm2-5-24b-a2b","2026-10-14T23:10:00Z","2026-10-14T23:12:56.999102Z","2026-10-14T23:12:56.999111Z",true,"agent",15,[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},"28e1fd65-7f26-41e8-b1ba-aa69d03dce84","上海AI实验室开源 Atria Dawn Preview:五项 Agent 基准居首","shanghai-ai-lab-atria-dawn-preview","2026-09-14T19:08:09+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"643addcf-6639-4c7b-b6b9-dc478bab4792","Nex-N2.5 三档齐发:1.6T MoE 权重开放,BrowseComp 92.6 超 Opus 5","nex-n2-5-open-weights-agent-family","2026-09-12T13:07:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"c05a1cb5-8f78-4729-812b-2e879284a51e","阶跃开源 600B MoE:输出低至每百万 2.7 美元","stepfun-step-5-preview-pareto-frontier","2026-10-15T05:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"63ae3eda-4994-4f9f-a812-489d9831a9ca","Clef-omni 开源:决策模型学会听音看片","clef-omni-multimodal-decision-model","2026-10-12T15:20:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"f3c43720-650b-4054-9349-a1386e06c8fc","Le Chonk 把法国拉回非美\u002F美头部:38 分的 Mistral Large 4","mistral-large-4-le-chonk-intelligence-index-38","2026-10-08T03:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"5ddba2a9-781e-4763-b1e0-1e20c6480391","Mistral Large 4:1万亿参数MoE,月底开源","mistral-large-4-1t-moe","2026-10-06T23:10:00+00:00"]