LFM2.5-24B-A2B 发布：为高速端侧 Agent 打造的推理模型

今天，我们发布 LFM2.5-24B-A2B：这是一款拥有 240 亿参数、每个 token 仅激活 20 亿参数的推理模型，专为快速、注重隐私的端侧 Agent 打造。它是 LFM2.5 系列中规模最大、也是最后发布的模型；该系列的模型规模现已跨越两个数量级，从 LFM2.5-230M 到 LFM2.5-24B-A2B。

在 8 项 Agent 与工具使用基准测试中，LFM2.5-24B-A2B 有 5 项表现与 Qwen3.6-35B-A3B 相当或更好；后者的总参数量和激活参数量均比前者多约 50%。在 LiveClaw 上，两者得分分别为 48.09 和 38.19；在 AutomationBench 上分别为 27.69 和 8.72。由于每个 token 仅激活 20 亿参数，该模型在端侧的运行速度比同类 MoE 模型最高快 1.7 倍：在 Apple M5 Max 上达到 184 token/秒，在单张 NVIDIA B300 GPU 上的输出吞吐量最高可达每秒 2.1 万 token。

基础模型（LFM2.5-24B-A2B-Base）和后训练模型（LFM2.5-24B-A2B）均开放权重，即日起可通过 Hugging Face 和我们的 Playground 获取。有关如何在本地运行和微调模型，请参阅文档。

相比 LFM2-24B-A2B 有哪些变化

LFM2.5-24B-A2B 在此前发布的 LFM2-24B-A2B 基础上构建，扩展了词表和上下文窗口，扩大了预训练规模，并引入大规模强化学习。与前代模型不同，LFM2.5-24B-A2B 是一款纯推理模型，在给出最终答案前会生成明确的思维链。我们采用这一策略，是为了利用其高速解码能力并提高答案质量。

架构。 模型保留了 40 层架构、每个 MoE 模块 64 个专家及 top-4 路由，以及门控短卷积与 GQA 相结合的混合架构。两项架构变化分别是将上下文窗口从 32,768 token 扩展至 131,072 token，以及通过原位分词器扩展将词表从 65K 增至 128K token，以改善非拉丁文字的分词效果。

预训练规模从 17T token 扩大至 33T token。

中期训练分为两个阶段。第一阶段在 32K 上下文长度下训练 2T token，使用经过筛选的高质量数据，并侧重推理、数学、代码、指令遵循和多语言文本。第二阶段在 131K 上下文长度下训练 300B token，侧重长上下文数据。两个阶段均包含工具调用和多轮交互轨迹，以帮助模型应对 Agent 工作负载。

后训练通过下图所示的五阶段流程，将 LFM2.5-24B-A2B-Base 转变为具备 Agent 能力的 LFM2.5-24B-A2B：（1）监督微调（SFT）；（2）教师模型专门化；（3）非 Agent 和（4）Agent 多领域在策略蒸馏（MOPD）；以及（5）Agent 强化学习（Agentic RL），即在 OpenClaw、Hermes Agent、Pi 和 OpenCode 等 Agent 运行框架中训练模型。

图 2：LFM2.5-24B-A2B 后训练流程

在教师模型专门化期间，LFM2.5-24B-A2B 出现了众所周知的奖励黑客行为：在代码仓库的 git 历史记录中，或在 GitHub、PyPI 和 Hugging Face 等平台上搜索最终答案。更令人惊讶的是，如下方示例所示，它甚至发现自己身处沙箱环境。

这种行为看起来像是涌现能力，但也可能是因为训练数据中包含类似模式。对此，我们加强了沙箱隔离；下文报告的所有结果均来自封闭的评估环境。

基准测试

我们通过 16 项基准测试评估了 LFM2.5-24B-A2B，涵盖工具使用与 Agent 工作流、知识、指令遵循、数学和代码。尽管它的总参数量和激活参数量均为对比模型中最小，但其表现可与总参数量和激活参数量多约 50% 的大型 MoE 模型（如 Qwen3.6-35B-A3B）相媲美，且经常超越后者。

基准测试	LFM2.5-24B-A2B	Qwen3.6-35B-A3B	Gemma-4-26B-A4B-it	Nemotron-3.5-Lightning (30B-A3B)
AutomationBench	27.69	8.72	5.46	2.61
BFCLv4	65.01	65.00	57.09	53.78
ToolSandbox	80.71	75.02	72.09	75.13
LiveClaw	48.09	38.19	41.25	37.72
PinchBench	83.18	89.93	77.12	84.20
τ³-Bench Banking	8.04	7.63	4.54	7.01
WebArena Verified	26.75	37.62	25.62	28.69
GPQA Diamond	74.42	83.69	74.90	72.42
AA-Omniscience	-24.90	-24.35	-59.39	-3.58
HLE	7.40	7.69	8.57	5.65
IFEval	96.75	93.06	92.00	94.19
IFBench	74.69	64.21	47.44	71.67
Multi-IF	84.38	81.53	82.19	82.34
MATH-500	96.04	96.73	95.66	96.40
SWE-Bench Verified (OpenCode)	51.00	66.20	25.80	53.40
SciCode	32.71	37.35	35.42	27.78

表 1：LFM2.5-24B-A2B 基准测试表现*。

在 8 项工具使用与 Agent 基准测试中，LFM2.5-24B-A2B 有 5 项领先或持平于竞争对手，分别是 AutomationBench、BFCLv4、ToolSandbox、LiveClaw 和 τ³-Bench Banking。其中它在 AutomationBench 上的优势最大，得分超过第二名的三倍。它还在全部 3 项指令遵循基准测试中领先。更大的模型在 GPQA Diamond、HLE 和 AA-Omniscience 等知识密集型基准测试，以及 SWE-Bench Verified（OpenCode）和 SciCode 等代码基准测试中仍保持优势。

这些结果表明，LFM2.5-24B-A2B 适合作为需要推理、调用工具并在设备上可靠运行的 Agent 的基础。对于知识密集型工作负载，我们建议将 LFM2.5-24B-A2B 与检索或网络搜索结合使用，以发挥其工具调用方面的优势。

随处实现快速推理

LFM2.5-24B-A2B 发布之初即获得推理生态系统的广泛支持：

• llama.cpp —— 提供 GGUF 检查点，支持高效边缘推理
• MLX —— 针对 Apple Silicon 优化的推理
• vLLM —— GPU 加速服务，满足生产环境吞吐量需求
• SGLang —— 高性能 GPU 服务框架
• ONNX —— 在多种加速器上实现跨平台推理

所有运行数据均通过 Pipette 收集，这是 Liquid AI 公开设备性能数据所使用的同一基准测试基础设施。

边缘推理。 得益于高效的 LFM2 架构，在参与对比的 MoE 模型中，LFM2.5-24B-A2B 读取提示词和生成回答的速度均最快。它在 Apple M5 Max 上的解码速度为 184 token/秒，在 AMD Ryzen AI Max+ 395 上为 114 token/秒，最高比第二快的模型快 1.7 倍，同时内存占用最低。

GPU 推理。 为展示 LFM 架构的竞争力，我们将其 GPU 推理吞吐量与其他类似规模的 MoE 模型进行比较。我们在单张 NVIDIA B300 GPU 上采用持续负载方式测量 GPU 吞吐量（总输出 token 数除以实际耗时）：在每个并发级别，持续保持目标数量的进行中请求，每当一个请求完成就立即补入新请求。

我们使用 SGLang 0.5.22 测试每个模型，输入长度为 1,024 token，输出长度最多为 256 token，使用 BF16 精度。LFM2.5-24B-A2B 是同规模模型中速度最快的，在高并发情况下每秒输出超过 2.1 万 token。

使用 LFM2.5-24B-A2B 运行本地 Agent

LFM2.5-24B-A2B 的规模、速度和能力使其适合用于在自有硬件上持续运行的 Agent。在下面的演示中，我们在一台搭载 M5 Max 的 MacBook Pro 上，将它置于 Hermes Agent 中运行；模型完全在设备本地进行规划、调用工具并完成一项真实任务，不调用任何云端 API。

视频 1：搭载 LFM2.5-24B-A2B 的本地 Hermes Agent 制作一份关于 Hugging Face 热门模型的晨间简报，并回答后续问题。

演示的重点并非单个工具，而是工具调度循环在消费级硬件上具有交互性：提问、提出方案、确认、运行、重复，每次调度均远低于一秒，同时保留完整审计记录，数据也不会离开设备。有关如何在本地部署模型并连接所选 Agent 运行框架，请参阅本地 Agent 指南。

开始使用

立即使用 LFM2.5-24B-A2B 开始构建；模型可通过 Hugging Face 获取。

如需通过 API 进行推理，也可以注册 Liquid Console，获得 2 美元的入门额度。模型价格为每 100 万输入 token 0.10 美元（缓存输入为 0.05 美元），每 100 万输出 token 1.00 美元。

通过 LFM2.5，我们正在实现随处运行 AI 的愿景。这些模型具有以下特点：

• 开放权重 —— 可下载、微调和部署。请参阅模型许可证。
• 开箱即快 —— 在 Apple、AMD、Qualcomm 和 Nvidia 硬件上原生支持 llama.cpp、MLX 和 vLLM。
• 完整的模型家族 —— 从用于定制的基础模型到专门的音频和视觉变体，同一架构覆盖多种使用场景。

我们迫不及待想看看你会构建出什么。

在 Hugging Face 下载

使用 LFM2.5-24B-A2B 在本地运行你选择的 Agent 框架

集成我们的 API

引用

引用时，请使用以下参考文献或 BibTeX：

Liquid AI, "LFM2.5-24B-A2B: Built for Fast On-Device Agents", Liquid AI Blog, Oct 2026.

bibtex @article{liquidAI202624BA2B, author = {Liquid AI}, title = {LFM2.5-24B-A2B: Built for Fast On-Device Agents}, journal = {Liquid AI Blog}, year = {2026}, note = {www.liquid.ai/blog/lfm2-5-24b-a2b}, }

*所有模型均使用 vLLM 和以下生成参数进行评估：

• AutomationBench： temperature = 0.6，top_p = 0.95，不限制最大输出 token 数（受模型上下文窗口限制）。
• BFCLv4： temperature = 0.001，最大输出 token 数 = 4096。
• ToolSandbox： temperature = 0，最大输出 token 数 = 1024；用户模拟器：GPT-4o-2024-08-06，temperature = 0，最大输出 token 数 = 1024。
• LiveClaw： temperature = 1.0（vLLM 默认值；Agent 不设置 temperature），最大输出 token 数 = 8192。
• τ³-Bench Banking： 使用官方用户模型（GPT-5.2，低推理强度）和默认 BM25 检索，temperature = 0，每轮最大输出 token 数 = 16384。
• WebArena： temperature = 0.6，top_p=0.95，max_turns=40（允许的浏览器操作次数），不限制最大输出 token 数（受模型上下文窗口限制）。
• 其他评估： temperature = 0.6，最大输出 token 数 = 32768。

Today, we release LFM2.5-24B-A2B , a 24B-parameter reasoning model with just 2B active parameters, built for fast, private agents on-device. It is the largest and final model in the LFM2.5 family, which now spans two orders of magnitude: from LFM2.5-230M to LFM2.5-24B-A2B.

LFM2.5-24B-A2B is competitive with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on five of eight agentic and tool-use benchmarks, scoring 48.09 on LiveClaw versus 38.19 and 27.69 on AutomationBench versus 8.72. Because only 2B parameters are active per token, it runs up to 1.7× faster than comparable MoEs on-device, with 184 tokens/s on an Apple M5 Max and up to 21K output tokens/s on a single NVIDIA B300 GPU.

The base (LFM2.5-24B-A2B-Base) and post-trained (LFM2.5-24B-A2B) models are open-weight and available today on Hugging Face and our Playground. Check out our docs on how to run and fine-tune it locally.

What changed since LFM2-24B-A2B

LFM2.5-24B-A2B builds on our previous LFM2-24B-A2B release, with an expanded vocabulary and context window, scaled-up pretraining, and large-scale reinforcement learning. Unlike its predecessor, LFM2.5-24B-A2B is a reasoning-only model that produces an explicit chain of thought before its final answer. We adopted this strategy to leverage its high decode speed and improve answer quality.

Architecture. It keeps the 40 layers, 64 experts per MoE block with top-4 routing, and the same hybrid of gated short convolutions and GQA. The two architectural changes are the context window extension from 32,768 to 131,072 tokens, and the in-place tokenizer expansion, which doubles the vocabulary from 65K to 128K tokens for better non-Latin tokenization.

Pre-training was scaled up from 17T tokens to 33T tokens.

Mid-training has two phases. The first runs 2T tokens at 32K context on curated, higher-quality data weighted toward reasoning, math, code, instruction following, and multilingual text. The second runs 300B tokens at 131K context, weighted toward long context data. Tool calls and multi-turn trajectories appear throughout both phases to prepare the model for agentic workloads.

Post-training

turns LFM2.5-24B-A2B-Base into the agentic LFM2.5-24B-A2B in a five-stage pipeline as illustrated below: (1) Supervised Fine-Tuning (SFT), (2) Teacher Specialization, (3) non-agentic and (4) agentic Multi-Domain On-Policy Distillation (MOPD), and (5) Agentic Reinforcement Learning (Agentic RL), training the model inside agent harnesses, such as

OpenClaw

,

Hermes Agent

,

Pi

, and

OpenCode

.

Figure 2: LFM2.5-24B-A2B Post-training recipe

During teacher specialization, LFM2.5-24B-A2B exhibited well-known reward-hacking behaviors: searching for the final answer in a repository's git history or on platforms like GitHub, PyPI, and Hugging Face. More surprisingly, it even figured out it was in a sandbox, as shown in the examples below.

This behavior looks emergent but might be explained by training data containing similar patterns. We tightened our sandboxes in response, and all reported results below come from sealed evaluation environments.

Benchmarks

We evaluated LFM2.5-24B-A2B across 16 benchmarks covering tool use and agentic workflows, knowledge, instruction following, math, and coding. Despite being the smallest model in total and active number of parameters, it is competitive with, and often outperforms, larger MoEs with ~50% more total and active parameters, such as Qwen3.6-35B-A3B.

Benchmark	LFM2.5-24B-A2B	Qwen3.6-35B-A3B	Gemma-4-26B-A4B-it	Nemotron-3.5-Lightning (30B-A3B)
AutomationBench	27.69	8.72	5.46	2.61
BFCLv4	65.01	65.00	57.09	53.78
ToolSandbox	80.71	75.02	72.09	75.13
LiveClaw	48.09	38.19	41.25	37.72
PinchBench	83.18	89.93	77.12	84.20
τ³-Bench Banking	8.04	7.63	4.54	7.01
WebArena Verified	26.75	37.62	25.62	28.69
GPQA Diamond	74.42	83.69	74.90	72.42
AA-Omniscience	-24.90	-24.35	-59.39	-3.58
HLE	7.40	7.69	8.57	5.65
IFEval	96.75	93.06	92.00	94.19
IFBench	74.69	64.21	47.44	71.67
Multi-IF	84.38	81.53	82.19	82.34
MATH-500	96.04	96.73	95.66	96.40
SWE-Bench Verified (OpenCode)	51.00	66.20	25.80	53.40
SciCode	32.71	37.35	35.42	27.78

Table 1: LFM2.5-24B-A2B benchmark performance*.

LFM2.5-24B-A2B leads or matches the competition on five of the eight tool-use and agentic benchmarks (AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking), with its widest margin on AutomationBench, where it scores more than three times the next-best model. It also leads on all three instruction-following benchmarks. Larger models keep an edge on knowledge-heavy benchmarks like GPQA Diamond, HLE, and AA-Omniscience, and on coding benchmarks like SWE-Bench Verified (OpenCode) and SciCode.

The results make LFM2.5-24B-A2B an ideal foundation for agents that need to reason, call tools, and run reliably on your device. For knowledge-heavy workloads, we recommend pairing LFM2.5-24B-A2B with retrieval or web search, which plays to its tool-calling strength.

Fast inference everywhere

LFM2.5-24B-A2B ships with day-one support across the inference ecosystem:

• llama.cpp — GGUF checkpoints for efficient edge inference
• MLX — Optimized inference for Apple Silicon
• vLLM — GPU-accelerated serving for production throughput
• SGLang — High-performance GPU serving framework
• ONNX — Cross-platform inference across diverse accelerators

All runs were collected on Pipette, the same benchmarking infrastructure behind Liquid AI's public device performance data.

Edge inference. Due to the efficient LFM2 architecture, LFM2.5-24B-A2B is the fastest of the compared MoEs at reading in prompts and generating answers. It decodes 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+ 395, up to 1.7x faster than the next-best model, while using the least memory.

GPU inference. To demonstrate the competitiveness of LFM architecture we compare the GPU inference throughput to other similarly sized MoE models. We measure GPU throughput (total output tokens / wall time) on a single NVIDIA B300 GPU using a sustained-load setting: at each concurrency level, we continuously maintain the target number of in-flight requests, replacing each completed request immediately.

We benchmark each model with SGLang 0.5.22, 1,024 input tokens, up to 256 output tokens, in BF16. LFM2.5-24B-A2B is the fastest model in its size class, reaching over 21K output tokens per second at high concurrency.

Run local agents with LFM2.5-24B-A2B

LFM2.5-24B-A2B’s size, speed, and capabilities make it a great choice for agents that run continuously on your own hardware. In the demo below, we run it inside Hermes Agent on an M5 Max MacBook Pro, where it plans, calls tools, and works through a real task entirely on-device, without any cloud API calls.

Video 1: Local Hermes Agent with LFM2.5-24B-A2B creating a morning briefing about trending models on Hugging Face and responding to follow up questions.

The point of the demo is not the individual tools. It is that the tool-dispatch loop feels interactive on consumer hardware: ask, propose, confirm, run, repeat, all in well under a second per dispatch, with full audit trails and your data never leaving the device. Check out our local agent guide for how to serve the model locally and connect it with the agent harness of your choice.

Get Started

Start building today with LFM2.5-24B-A2B, available on Hugging Face.

For API inference, you can also sign up for Liquid Console where you'll receive $2 in credit to get started. The model is priced at $.10/1M input tokens (at $.05 for cached input) and $1.00/1M output tokens.

With LFM2.5, we're delivering on our vision of AI that runs anywhere. These models are:

• Open-weight — Download, fine-tune, and deploy. See our model license.
• Fast from day one — Native support for llama.cpp, MLX, and vLLM across Apple, AMD, Qualcomm, and Nvidia hardware
• A complete family — From base models for customization to specialized audio and vision variants, one architecture covers diverse use cases

We can't wait to see what you build.

Download on Hugging FaceRun your harness of choice locally with LFM2.5-24B-A2BIntegrate with our API

Citation

For citations, please use the following reference or BibTeX:

Liquid AI, "LFM2.5-24B-A2B: Built for Fast On-Device Agents", Liquid AI Blog, Oct 2026.

@article{liquidAI202624BA2B,
  author = {Liquid AI},
  title = {LFM2.5-24B-A2B: Built for Fast On-Device Agents},
  journal = {Liquid AI Blog},
  year = {2026},
  note = {www.liquid.ai/blog/lfm2-5-24b-a2b},
}

*All the models were evaluated with vLLM and the following generation parameters:

• AutomationBench: temperature = 0.6, top_p = 0.95, no max output tokens (bounded by the model's context window).
• BFCLv4 : temperature = 0.001, max output tokens = 4096.
• ToolSandbox : temperature = 0, max output tokens = 1024; user simulator: GPT-4o-2024-08-06, temperature = 0, max output tokens = 1024.
• LiveClaw : temperature = 1.0 (vLLM default; the agent sets no temperature), max output tokens = 8192.
• τ³-Bench Banking : official user model (GPT-5.2, low reasoning effort) with default BM25 retrieval, temperature = 0, max output tokens = 16384 per turn.
• WebArena: temperature = 0.6, top_p=0.95, max_turns=40 (no of browser actions allowed), no max output tokens (bounded by the model's context window))
• Other evals : temperature = 0.6, max output tokens = 32768.