[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-lfm2-5-2-6b-on-device-agent":3,"news-related-5878a668-282c-4b88-b2b8-7eef40b7938c":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"5878a668-282c-4b88-b2b8-7eef40b7938c","LFM2.5-2.6B：2.5GB 内存跑本机 Agent 220 tok\u002Fs","Liquid AI 发布 LFM2.5-2.6B,2.6B 参数 dense 模型专门面向 agent 工作流,预训练数据 ~34T tokens,词表扩到 128K 支持 128K 上下文,在 BFCLv4、ToolSandbox、IFBench、Multi-IF、IFStruct 等 agent\u002Ftool-use benchmark 上接近甚至超过 9B 量级的 Qwen3.5-9B;M5 Max 跑出 220 tokens\u002Fs、手机 30 tokens\u002Fs、内存 \u003C 2.5GB,直接对接 Hermes Agent、OpenClaw、Pi 等本地 agent harness。","## LFM2.5-2.6B:把「本机 Agent」跑成 220 tokens\u002Fs,2.5GB 内存硬扛 Hermes Agent \u002F OpenClaw\n\n边缘 LLM 在过去半年里换了主角:从「能不能在手机上聊天」变成「能不能在手机上跑 Agent」。Liquid AI 在 2026 年 8 月 4 日放出的 LFM2.5-2.6B,正是为这条新赛道定制的旗舰:2.6B 参数 dense 模型,128K 上下文,词表扩到 128K,总内存占用压在 2.5GB 以下,但所有 agentic benchmark 都在和 9B 量级的 Qwen3.5-9B 掰手腕。\n\n### 模型规格与训练数据\n\nLFM2.5-2.6B 是一款面向 agentic 工作流的 dense 模型,总参数 2.6B,预训练数据约 34T tokens。Liquid AI 这次没走 MoE 路线(8B-A1B 才用 MoE),2.6B 选 dense 的目的是让权重在边缘硬件上更易调度。词表从 LFM2 时代的约 64K 翻倍到 128K,方法是「在原地续接 BPE,子词均值初始化」,而不是重新训模型;mid-training 还专门跑了一个 128K context-extension 阶段,以便应对 agent 工作流里的长输入。\n\n后训练走的是一条四阶段管线。Liquid AI 把这套管线讲得很细:\n\n1. **Supervised Fine-Tuning(SFT)**:两段 SFT,第一段广覆盖,第二段重点塑形 agent 任务、推理和工具调用;SFT 训练数据量约为 LFM2.5-8B-A1B 的 7 倍。\n2. **Teacher Specialization**:从同一个 SFT checkpoint 分支出多个领域专家,每个专家在自己的领域(reasoning、knowledge、code、tool use、long context 等)用 RLVR(可验证奖励的强化学习)深耕。\n3. **Multi-Domain On-Policy Distillation(MOPD)**:学生模型在自己的 policy 下 rollout,不同 prompt 按领域路由到对应教师,教师提供 token 级监督——既保持学生分布稳定,又能让多个领域能力在一个模型里收敛。\n4. **Agentic Reinforcement Learning(Agentic RL)**:在真实 agent 环境里多轮 RL,用 GRPO 优化;训练里直接调用 Hermes Agent、OpenClaw 等 harness,让模型「在它将来要跑的沙盒里」学。\n\n值得注意的工程细节是,Harness Proxy 把 agent harness 当成 black box,透明地捕获 token 级轨迹来重建 RL 样本,这意味着 LFM2.5-2.6B 学到的不是「通用 tool use」,而是「在 Hermes Agent \u002F OpenClaw \u002F Pi 这一类 harness 里稳定工作」。\n\n### Benchmark:小四倍的体量,工具调用打到接近 9B\n\n官方在 BFCLv4、ToolSandbox、PinchBench、τ³-Bench、Claw-Eval、BrowseComp+、AIME25、IFBench、Multi-IF、IFStruct、AA-Omniscience、LiveCodeBenchv6 这一长串 benchmark 上做了对比。横向对手是 gemma-4-E2B-it(5.1B)、gemma-4-E4B-it(8B)、Qwen3.5-4B(4.7B)、Qwen3.5-9B(9.7B)——也就是体量上几乎都是 LFM2.5-2.6B 的 1.5× 到 4×。\n\nLFM2.5-2.6B 在所有指令遵循 benchmark 上全部领先:Multi-IF 拿到 80.07,而 Qwen3.5-9B 只有 62.55;IFStruct 拿到 85.49,Qwen3.5-9B 78.50;AA-Omniscience-Public 拿到了 -29.50 的成绩,远好于 Qwen3.5-9B 的 -50.43。工具调用维度,BFCLv4 是 56.88(略输 Qwen3.5-9B 的 60.13),ToolSandbox 77.83(超过 Qwen3.5-9B 的 76.44),PinchBench 68.22(略输 Qwen3.5-9B 的 71.45)。AIME25 数学 51.87,输给 Qwen3.5-9B 的 56.07 但远好于 Gemma-4-E4B 的 34.27。\n\n一句话总结官方在博客里给出的判断:LFM2.5-2.6B 在所有指令遵循 benchmark 和几乎所有工具调用 benchmark 上领先,只在数学和代码生成上稍微输给 9B 量级。\n\n### 推理速度:手机能跑、H100 能打 1.3B tokens\u002F天\n\n部署侧 Liquid AI 给出的数字非常具体:\n\n- **M5 Max(CPU)**:220 tokens\u002Fs,内存 \u003C 2.5GB。\n- **Ryzen AI Max+ 395(CPU)**:113 tokens\u002Fs。\n- **手机**:30 tokens\u002Fs,跑得动完整的 multi-step agent 流程。\n- **NVIDIA H100 SXM5**:SGLang 0.5.16、1,024 input \u002F 256 output、BF16 测下来,高并发下接近 15K output tokens\u002Fs,折算下来单卡每天能吐约 1.3B tokens。\n\nday-one 支持的推理栈包括 llama.cpp(GGUF)、MLX(Apple Silicon)、vLLM、SGLang、ONNX,覆盖 Apple、AMD、Qualcomm、NVIDIA 全平台。Hugging Face 上开源的权重包含 base(LFM2.5-2.6B-Base)和 post-trained(LFM2.5-2.6B)两个版本。\n\n### 为什么「本机 Agent」这件事现在变得重要\n\n把 Agent 完全放在本机跑,核心价值不是「能跑」,而是「去掉每 token 的边际成本」。Liquid AI 在博客里把这个动机写得很直白:当 token spend 不再是约束时,agent 可以被高度并行化,后台跑 background task,消耗百万 tokens 也不会让账单跳——这对开发者写 agent 的方式是一个根本性的变化。\n\n而把 LFM2.5-2.6B 选为「旗舰」也很合理:LFM2-8B-A1B(1.5B 激活 MoE)在 6 月已经给边缘 LLM 抬过一次天花板,LFM2.5-VL-450M 在 6 月把边缘 VLM 拉到亚秒级。这次 2.6B dense 把目标定在「能稳定接 Hermes Agent \u002F OpenClaw \u002F Pi 这类 harness」,等于是在「8B 边缘 MoE」和「350M \u002F 450M 极小模型」之间,填了一个「刚刚好能跑完整 agent」的甜点位。\n\n### 评论与局限\n\n从工程视角,LFM2.5-2.6B 的亮点不在参数,而在「把 agent 训练和 harness 训练绑在了一起」。通常的 LLM 训练只到 tool use 这一层,agent 行为是在 harness 那一层被 heuristic + prompt 堆出来的;Liquid AI 反过来,在 RL 阶段直接把 Hermes Agent \u002F OpenClaw 当训练环境,等于把「系统提示、工具协议、多轮交互模式」都写进了模型权重。这种做法对自家 harness 友好,但也带来一个开放问题:模型在其他 harness 上的迁移性会不会打折?官方没有给出跨 harness 的对照数据。\n\n另一个值得关注的限制是,BrowseComp+(OpenClaw)和 PinchBench 都只在 OpenClaw 上跑——这两个 benchmark 与「任务型 multi-step agent」高度相关,但评测本身就跑在 OpenClaw 里,没法完全区分「模型能力」与「harness 适配度」。对想复现或独立验证的开发者,这一点需要小心。\n\n最后一点:代码生成仍然是 9B 模型的领域。LiveCodeBenchv6 上 LFM2.5-2.6B 是 59.41,Qwen3.5-9B 是 69.86——差距是 10 个点。官方博客自己也很坦诚:「for more complex agentic tasks or coding-heavy workloads, larger models may still be a better fit」。2.6B 拿下的是「跑得起来、跑得便宜、跑得私密」,而不是「拿下所有最强模型」。\n\n### 结尾\n\nLFM2.5-2.6B 在 BFCL、ToolSandbox、IFBench、Multi-IF 这一类「agent 真实工作能力」的 benchmark 上,用自己的体量给出了接近甚至超过 9B 模型的成绩;M5 Max 上 220 tokens\u002Fs、内存 \u003C 2.5GB,把这个成绩塞进了边缘设备。Hermes Agent、OpenClaw、Pi 的用户可以在 Hugging Face 直接下载权重,接进现成的 harness 试跑。\n\n「本机 Agent」从口号到可量化产品的距离,Liquid AI 用这个 2.6B 模型量出来了——剩下的只是开发者愿不愿意把云上的 token 账单砍掉,换成本地的小时耗电。\n\n参考链接:\n- Liquid AI 官方博客:[LFM2.5-2.6B: Deploy Agents Everywhere](https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm2-5-2-6b)\n- Hugging Face 模型页:[LiquidAI\u002FLFM2.5-2.6B](https:\u002F\u002Fhuggingface.co\u002FLiquidAI\u002FLFM2.5-2.6B)\n- Hugging Face 博客:[Deploy local agents everywhere with LFM2.5-2.6B](https:\u002F\u002Fhuggingface.co\u002Fblog\u002FLiquidAI\u002Flfm2-5-2-6b)\n- Liquid AI 文档:[LFM2.5-2.6B - Liquid Docs](https:\u002F\u002Fdocs.liquid.ai\u002Flfm\u002Fmodels\u002Flfm25-2.6b)","https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm2-5-2-6b","511bb1e6-a31f-4dc1-929b-9a7582e67447",[11,15,18,21,24],{"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},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":25,"name":26,"slug":26,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"0364b819-a989-445c-bc79-d3072313ce1c","en","LFM2.5-2.6B: local agents at 220 tok\u002Fs in 2.5GB","Liquid AI released LFM2.5-2.6B, a 2.6B-parameter dense model purpose-built for agentic workloads. Pre-trained on ~34T tokens, with vocabulary expanded to 128K and a 128K context window, it matches or beats the 9B-class Qwen3.5-9B on BFCLv4, ToolSandbox, IFBench, Multi-IF, and IFStruct. On an M5 Max it runs at 220 tokens\u002Fs under 2.5GB of memory and drops to 30 tokens\u002Fs on a phone, plugging directly into Hermes Agent, OpenClaw, and Pi.","## LFM2.5-2.6B: 220 tokens\u002Fs local agents under 2.5GB, wired into Hermes Agent and OpenClaw\n\nThe center of gravity in edge LLMs has shifted over the last six months. The question is no longer \"can this run a chat on a phone\" but \"can this drive an agent on a phone.\" Liquid AI's LFM2.5-2.6B, released on August 4, 2026, is built exactly for that new target: a 2.6B-parameter dense model, 128K context, a 128K-token vocabulary, and a memory footprint held under 2.5GB — while trading punches with the 9B-class Qwen3.5-9B on every agentic benchmark.\n\n### Spec and training data\n\nLFM2.5-2.6B is a dense model aimed squarely at agentic workloads: 2.6B total parameters, pre-trained on roughly 34T tokens. Liquid AI stayed dense here (the 8B-A1B variant was the MoE one); the point of going dense is that weights are easier to schedule on edge hardware. The vocabulary is doubled from the LFM2 era's ~64K to 128K by *extending* the existing BPE tokenizer in place with subword-mean initialization rather than re-training the model. Mid-training includes a dedicated 128K context-extension phase so the model can absorb the long inputs that agentic workflows need.\n\nPost-training runs through a four-stage pipeline. Liquid AI is unusually explicit about it:\n\n1. **Supervised Fine-Tuning (SFT).** Two SFT stages: broad coverage first, then a targeted shaping pass on agent tasks, reasoning, and tool use. The total SFT mix is roughly 7× the size of the one used for LFM2.5-8B-A1B.\n2. **Teacher Specialization.** A set of domain experts branch from the same SFT checkpoint and each deep-dives its area (reasoning, knowledge, code, tool use, long context) with RLVR.\n3. **Multi-Domain On-Policy Distillation (MOPD).** The student model rolls out under its own policy; per-domain teachers are routed to give token-level supervision. This is closer to the student's distribution than off-policy distillation and lets multiple domain capabilities converge inside a single model.\n4. **Agentic Reinforcement Learning (Agentic RL).** Multi-turn agentic RL inside real harnesses, optimized with GRPO and an outcome reward that combines an LLM-as-judge rubric, programmatic checks, and a hard safety gate. Training directly inside Hermes Agent, OpenClaw, and other harnesses means the model learns their tool APIs, system prompts, and interaction patterns.\n\nA piece of engineering worth flagging: the Harness Proxy treats agent harnesses as black boxes and transparently captures token-level trajectories to reconstruct RL samples. In practice, LFM2.5-2.6B is not just learning \"general tool use\" — it is being trained to behave reliably *inside Hermes Agent \u002F OpenClaw \u002F Pi*.\n\n### Benchmarks: a quarter of the size, near 9B-class tool use\n\nLiquid AI published results on BFCLv4, ToolSandbox, PinchBench, τ³-Bench, Claw-Eval, BrowseComp+, AIME25, IFBench, Multi-IF, IFStruct, AA-Omniscience, and LiveCodeBenchv6. The comparators are gemma-4-E2B-it (5.1B), gemma-4-E4B-it (8B), Qwen3.5-4B (4.7B), and Qwen3.5-9B (9.7B) — between 1.5× and 4× the size of LFM2.5-2.6B.\n\nLFM2.5-2.6B wins every instruction-following benchmark: Multi-IF 80.07 vs. Qwen3.5-9B's 62.55; IFStruct 85.49 vs. 78.50; AA-Omniscience-Public -29.50, a wide margin over Qwen3.5-9B's -50.43. On tool use, BFCLv4 is 56.88 (a touch below Qwen3.5-9B's 60.13), ToolSandbox is 77.83 (above 76.44), and PinchBench is 68.22 (below 71.45). AIME25 lands at 51.87 — losing to Qwen3.5-9B's 56.07 but well clear of Gemma-4-E4B's 34.27.\n\nThe blog's own summary: LFM2.5-2.6B leads on every instruction-following benchmark and on almost every tool-use benchmark, only losing on math and code generation against the 9B.\n\n### Inference speed: a phone can run it; an H100 can spit out 1.3B tokens\u002Fday\n\nLiquid AI's deployment numbers are concrete:\n\n- **M5 Max (CPU):** 220 tokens\u002Fs, \u003C 2.5 GB of memory.\n- **Ryzen AI Max+ 395 (CPU):** 113 tokens\u002Fs.\n- **Smartphone:** 30 tokens\u002Fs, enough for a complete multi-step agent loop.\n- **NVIDIA H100 SXM5:** SGLang 0.5.16, 1,024 input \u002F 256 output, BF16, sustaining high concurrency. LFM2.5-2.6B reaches nearly 15K output tokens\u002Fs, i.e. about 1.3B tokens per day on a single card.\n\nDay-one inference stack coverage includes llama.cpp (GGUF), MLX (Apple Silicon), vLLM, SGLang, and ONNX — Apple, AMD, Qualcomm, and NVIDIA. The open weights on Hugging Face ship in two flavors: base (LFM2.5-2.6B-Base) and post-trained (LFM2.5-2.6B).\n\n### Why \"local agents\" matters right now\n\nRunning an agent entirely on-device is not about feasibility; it's about removing the per-token marginal cost. Liquid AI states the motivation bluntly: when token spend is no longer a constraint, agents can be massively parallelized on local hardware, run as background tasks, burn through millions of tokens with no marginal cost. That changes how developers build agents.\n\nChoosing LFM2.5-2.6B as the \"flagship\" of the LFM2.5 family makes sense given that lineup. LFM2-8B-A1B (1.5B-active MoE) lifted the ceiling for edge LLMs in June. LFM2.5-VL-450M pushed edge VLMs to sub-second in June. The 2.6B dense model fills a different gap: the sweet spot where a full multi-step agent actually runs reliably, not the smallest possible chat model and not the heaviest edge MoE.\n\n### Commentary and caveats\n\nFrom an engineering angle, the headline is not the parameter count — it's that agent training and harness training are bundled together. Most LLM training stops at the \"tool use\" layer; agent behavior is assembled on top by the harness's prompts and heuristics. Liquid AI runs RL *inside* Hermes Agent and OpenClaw, which means the model's weights absorb the system prompt, the tool protocol, and the multi-turn interaction pattern. That is friendly to the targeted harnesses, but it raises an open question: does the model transfer cleanly to other harnesses? Liquid AI does not publish cross-harness results.\n\nAnother caveat: BrowseComp+ (OpenClaw) and PinchBench are reported only against OpenClaw. These benchmarks are tightly coupled to task-style multi-step agent behavior, but evaluating inside the harness itself blurs the line between *model capability* and *harness fit*. Reproducers and independent reviewers should keep that in mind.\n\nFinally, code generation is still the 9B models' territory. LiveCodeBenchv6: LFM2.5-2.6B 59.41 vs. Qwen3.5-9B 69.86 — a 10-point gap. Liquid AI's own blog is candid: \"for more complex agentic tasks or coding-heavy workloads, larger models may still be a better fit.\" What the 2.6B buys you is \"runs, runs cheaply, runs privately\" — not \"wins every benchmark.\"\n\n### Closing\n\nOn BFCL, ToolSandbox, IFBench, and Multi-IF — the benchmarks that most closely track real agent capability — LFM2.5-2.6B matches or beats 9B-class models from a quarter of the size. On an M5 Max it runs at 220 tokens\u002Fs under 2.5 GB of memory. Hermes Agent, OpenClaw, and Pi users can pull the weights directly from Hugging Face and wire them into existing harnesses today.\n\nLiquid AI has put a number on the gap between \"local agents\" as a slogan and as a measurable product. What remains is whether developers are willing to swap their cloud token bills for local kilowatt-hours.\n\nReferences:\n- Liquid AI blog: [LFM2.5-2.6B: Deploy Agents Everywhere](https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Flfm2-5-2-6b)\n- Hugging Face model card: [LiquidAI\u002FLFM2.5-2.6B](https:\u002F\u002Fhuggingface.co\u002FLiquidAI\u002FLFM2.5-2.6B)\n- Hugging Face blog: [Deploy local agents everywhere with LFM2.5-2.6B](https:\u002F\u002Fhuggingface.co\u002Fblog\u002FLiquidAI\u002Flfm2-5-2-6b)\n- Liquid AI docs: [LFM2.5-2.6B - Liquid Docs](https:\u002F\u002Fdocs.liquid.ai\u002Flfm\u002Fmodels\u002Flfm25-2.6b)","lfm2-5-2-6b-on-device-agent","2026-08-11T00:00:00Z","2026-08-11T14:06:30.877711Z","2026-08-19T01:48:03.231362Z",true,"agent",151,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"f26ace13-9c96-47ea-a528-b6682a22aa1e","Apodex 1.1 把推理搬进真实执行:PIVOT-RL 定位关键决策点,35B mini 开源","apodex-1-1-agentic-execution-pivot-rl","2026-08-25T14:30:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"b4754043-6b19-499f-8459-f8fc786f4d80","Pokee-Isaac 28B 把 10M 上下文塞进客户边界:28B 参数在 RULER 10M 上 93.3%","pokee-isaac-28b-10m-context","2026-08-20T14:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"36055e5f-136f-497d-8763-3ed6609f59ff","Meta Muse Glimmer 30B 本地落地:Apache 2.0 的开源智能体,把 Agent 装进 24GB 显存","meta-muse-glimmer-30b-local-agent-apache2-r2","2026-08-19T03:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"79c1684f-f61d-4799-b3d0-6450c4ad10e8","Muse Glimmer:Meta 把 30B 「常驻本地的智能体」开源,把 Agent 拉到笔记本里 7×24 跑","muse-glimmer-30b-open-agentic-local","2026-08-10T00:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"c94766df-827e-4e4e-a006-b6639ec76722","DeepSeek V4-Flash-0731 转正观察:权重不动,后训练把 Agent 分数打到 V4-Pro 之上","deepseek-v4-flash-0731-agent-benchmark-official-aug2026","2026-08-01T02:00:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"15987a0e-bc06-4f21-9608-264da02d0e6c","AutoMem 让 32B 开源模型在长程任务上追平 Claude Opus 4.5","automem-stanford-32b-long-horizon","2026-07-23T12:10:00+00:00"]