Liquid AI Releases LFM2.5-24B-A2B Reasoning Model for On-Device Agents

Liquid AI Releases LFM2.5-24B-A2B Reasoning Model for On-Device Agents Liquid AI has released LFM2.5-24B-A2B, an open-weight 24-billion-parameter reasoning model featuring 2 billion active parameters per token, designed for fast and private on-device agent execution. Retains a 40-layer MoE architecture with top-4 routing among 64 experts, expanding context length to 131K tokens and vocabulary to 128K tokens. Pre-training scaled to roughly 33 trillion tokens, followed by 2.3 trillion mid-training tokens and a five-stage post-training pipeline ending in agentic RL. Outperformed or matched competing MoE models across multiple agentic benchmarks, scoring 27.69 on AutomationBench and 48.09 on LiveClaw. Delivers edge decoding speeds of 184 tokens/s on Apple M5 Max and up to 21,000 output tokens/s on a single NVIDIA B300 GPU. Architecture and Training Pipeline LFM2.5-24B-A2B builds upon LFM2-24B-A2B using 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid structure of gated short convolutions and GQA. Its context window is expanded to 131K tokens alongside a doubled vocabulary of 128K tokens. Pre-training was expanded from ~17T to ~33T tokens, succeeded by mid-training comprising 2T tokens at 32K context and 300B tokens at 131K context. Post-training is structured in five stages: supervised fine-tuning, teacher specialization, non-agentic multi-domain on-policy distillation (MOPD), agentic MOPD, and agentic reinforcement learning across harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode. Benchmark Evaluation Tested on 16 benchmarks against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning, LFM2.5-24B-A2B matched or led on five of eight tool-use and agentic tasks. It posted 27.69 on AutomationBench, 65.01 on BFCLv4, 80.71 on ToolSandbox, 48.09 on LiveClaw, and 8.04 on τ³-Bench Banking, while also leading all three instruction-following evaluations. Larger models maintained an edge on coding benchmarks like SWE-Bench Verified (51.00) and knowledge-heavy assessments including GPQA Diamond (74.42), indicating benefits from pairing LFM2.5-24B-A2B with retrieval or web search. Inference and Deployment The model features day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. On edge hardware, it achieves 184 tokens/s on an Apple M5 Max and 114 tokens/s on a Ryzen AI Max+ 395, while reaching over 21,000 output tokens/s on an NVIDIA B300 GPU under high concurrency. Both the base and tuned models are available on Hugging Face, support transformers>=5.0.0, and can be integrated locally with agent harnesses using OpenAI-compatible endpoints.

Liquid AI 推出端侧推理智能体模型 LFM2.5-24B-A2B Liquid AI 发布了面向端侧快速、私密智能体运行的推理模型 LFM2.5-24B-A2B。该模型总参数量为 24B，每个 token 仅激活 2B 参数，支持工具调用与多步智能体任务，并在生成最终答案前进行显式链式思考。 采用 40 层、64 专家（top-4 路由）架构，上下文窗口拓展至 131K，词表扩大到 128K。 经过约 33T token 预训练、中训练以及包含 SFT、教师特化、双阶段 MOPD 和智能体强化学习的五阶段后训练。 在 AutomationBench、LiveClaw 等工具调用与指令遵循基准上领先或持平更大模型，在 M5 Max 上推理速度达 184 tok/s。 架构设计与五阶段后训练 LFM2.5-24B-A2B 沿用了门控短卷积与 GQA 混合的 MoE 架构，包含 40 层并在每个 MoE 块设 64 个专家（top-4 路由）。模型上下文窗口从 32K 扩至 131K，词表从 65K 翻倍至 128K。预训练规模提升至约 33T tokens，随后在中训练加入 2T（32K 上下文）和 300B（131K 上下文）数据。后训练则分为监督微调（SFT）、教师特化、非智能体任务多域策略蒸馏（MOPD）、智能体任务 MOPD，以及在 OpenClaw、Hermes Agent 等智能体运行框架中的强化学习训练五个阶段。 基准测试表现与适用场景 在覆盖工具使用、智能体流程、知识、数学及代码的 16 项基准测试中，LFM2.5-24B-A2B 在 8 项工具使用测试中的 5 项（包括 AutomationBench、BFCLv4、LiveClaw 等）领先或追平总参数和激活参数高出约 50% 的模型；在 AutomationBench 上得分达到次优模型的 3 倍以上，并领跑全部 3 项指令遵循测试。但在 GPQA Diamond 等知识密集型和 SWE-Bench 等代码基准上弱于更大模型，因此在重知识型工作流中建议搭配检索或网络搜索。 端侧推理性能与生态部署 该模型开箱支持 llama.cpp、MLX、vLLM、SGLang 和 ONNX。端侧推理方面，其在 Apple M5 Max 上的解码速度为 184 tok/s，在 Ryzen AI Max+ 395 上达 114 tok/s；在单块 NVIDIA B300 GPU 上的高并发输出可超过 21K tok/s。开源权重已在 Hugging Face 提供，可通过兼容 OpenAI 标准的本地后端服务接入 Hermes Agent 等智能体框架直接调用。

LFM2.5-24B-A2B is a 24B-parameter reasoning model with just 2B active parameters, built for fast, private agents on-device. It thinks before it answers, calls tools, and works through multi-step tasks inside popular agent harnesses. Because only 2B parameters are active per token, each reasoning token is cheap, so you can run agents continuously on your own hardware. Best-in-class agent: Competitive with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use tasks, scoring 48.09 on LiveClaw versus 38.19. Reasoning at MoE speed: Always thinks before it answers. With only 2B active parameters per token, each reasoning token is cheap. Efficient inference: 184 tok/s on an Apple M5 Max and up to 21K output tok/s on a single NVIDIA B300 GPU. How we built a reasoning agent with 2B active parameters LFM2.5-24B-A2B builds on LFM2-24B-A2B and keeps its architecture: 40 layers, 64 experts per MoE block with top-4 routing, and the same hybrid of gated short convolutions and GQA. The context window grows from 32K to 131K tokens, and an in-place tokenizer expansion doubles the vocabulary from 65K to 128K tokens for better non-Latin tokenization. Unlike its predecessor, it is a reasoning-only model that produces an explicit chain of thought before its final answer. Pre-training was scaled up from ~17T to ~33T tokens. Mid-training then adds 2T tokens at 32K context, weighted toward reasoning, math, code, instruction following, and multilingual text, and 300B tokens at 131K context, weighted toward long-context data. Tool calls and multi-turn trajectories appear throughout both phases. Post-training turns the base model into an agent in five stages: Supervised fine-tuning (SFT) Teacher specialization Multi-domain on-policy distillation (MOPD), non-agentic: distill the specialist teachers on non-agentic tasks. MOPD, agentic: a second distillation stage focused on agentic tasks. Agentic reinforcement learning (Agentic RL): train the model inside agent harnesses, such as OpenClaw , Hermes Agent , Pi , and OpenCode . Benchmark results We evaluated LFM2.5-24B-A2B against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning on 16 benchmarks covering tool use and agentic workflows, knowledge, instruction following, math, and coding. It is the smallest model in the group, yet it is competitive with MoEs with ~50% more total and active parameters. 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 *All the models were evaluated with vLLM, using each benchmark's recommended generation parameters. For your app, the strengths are tool use and instruction following. 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), and on AutomationBench 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, so pair LFM2.5-24B-A2B with retrieval or web search for knowledge-heavy workloads. These results make LFM2.5-24B-A2B a strong foundation for agents that need to reason, call tools, and run reliably on your device. Inference speed on CPU and GPU LFM2.5-24B-A2B ships with day-one support across the inference ecosystem, including llama.cpp, MLX, vLLM, SGLang, and ONNX. Edge inference. LFM2.5-24B-A2B is the fastest of the compared MoEs, decoding 184 tokens/s on an M5 Max and 114 tokens/s on a Ryzen AI Max+ 395. That is up to 1.7x faster than the next-best model, while using the least memory. GPU inference. LFM2.5-24B-A2B is the fastest model in its size class, reaching over 21K output tokens per second at high concurrency on a single B300. How to use LFM2.5-24B-A2B Reach for LFM2.5-24B-A2B when you need agents that reason, call tools, and run reliably on your device. Install the latest version of transformers (compatible with transformers>=5.0.0 ): pip install -U transformers Then load and run the model: from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer model_id = "LiquidAI/LFM2.5-24B-A2B" model = AutoModelForCausalLM.from_pretrained( model_id, device_map= "auto" , dtype= "bfloat16" , # attn_implementation="flash_attention_2" <- uncomment on compatible GPU ) tokenizer = AutoTokenizer.from_pretrained(model_id) streamer = TextStreamer(tokenizer, skip_prompt= True , skip_special_tokens= True ) prompt = "What is C. elegans?" input_ids = tokenizer.apply_chat_template( [{ "role" : "user" , "content" : prompt}], add_generation_prompt= True , return_tensors= "pt" , tokenize= True , )[ "input_ids" ].to(model.device) output = model.generate( input_ids, do_sample= True , temperature= 0.1 , top_k= 50 , repetition_penalty= 1.05 , max_new_tokens= 512 , streamer=streamer, ) Use it in your agent harness Serve LFM2.5-24B-A2B locally behind any OpenAI-compatible backend, then point your agent harness at it. For example, with Hermes Agent: hermes config set model.provider custom hermes config set model.base_url http://localhost:8080/v1 hermes config set model.default LFM2.5-24B-A2B hermes config set model.context_length 131072 hermes config set model.api_mode chat_completions hermes config set agent.tool_use_enforcement true The port depends on your serving backend: llama.cpp and MLX use 8080 , vLLM uses 8000 , SGLang uses 30000 , and LM Studio uses 1234 . See the model card for OpenClaw and Pi configs, and our guide for the full setup. Check out our release blog for video examples. Get Started with LFM2.5-24B-A2B Both LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base are open-weight and available on Hugging Face today: Download: LFM2.5-24B-A2B-Base and LFM2.5-24B-A2B on Hugging Face. Use in your harness: follow our guide on how to run a local agent, like OpenClaw, Hermes Agent, and Pi. Fine-tune: adapt LFM2.5-24B-A2B to your task with our fine-tuning tutorials . We can't wait to see what you build. Citation If you use this work, please cite the release blog: @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}, }