Liquid AI Releases LFM2.5-2.6B for On-Device Agent Deployment

Liquid AI Releases LFM2.5-2.6B for On-Device Agent Deployment Liquid AI has launched LFM2.5-2.6B, a lightweight 2.6-billion-parameter language model designed for local agent workflows and tool calling on edge devices, maintaining competitive performance against models up to four times its size. Pre-trained on approximately 34 trillion tokens, with mid-training extending context length to 128K. Post-training relies on four stages: two rounds of agentic SFT, teacher specialization, multi-domain on-policy distillation, and Agentic Reinforcement Learning inside real agent harnesses. Consistently leads in instruction-following and tool-use benchmarks among compared models, though larger models retain an edge in coding. Decodes at 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen CPU in under 2.5 GB of memory. Training and Agentic RL Pipeline LFM2.5-2.6B was pre-trained on roughly 34 trillion tokens and expanded to a 128K context window during mid-training. Its post-training regime involves four stages: dual-round supervised fine-tuning oriented around tool use and harness trajectories, domain-specific teacher models, multi-domain on-policy distillation, and Agentic Reinforcement Learning. The RL system isolates training, rollouts, and environment execution, using a Harness Proxy to capture token-level trajectories from black-box agent harnesses like OpenClaw and Hermes Agent without modifying them. Benchmark Evaluation When evaluated against larger alternatives including Gemma-4 and Qwen3.5 variants up to 9.7B parameters, LFM2.5-2.6B achieved top scores in instruction-following tests like IFBench, Multi-IF, and IFStruct. It also led across most tool-use benchmarks, trailing only Qwen3.5-9B on BFCLv4. On general agentic benchmarks, it outperformed the Gemma models and matched Qwen models, while larger models maintained an advantage in coding tasks such as LiveCodeBenchv6. Inference Efficiency and Availability Operating in under 2.5 GB of memory, LFM2.5-2.6B reaches decode speeds of 220 tokens per second on an Apple M5 Max, 113 tokens per second on an AMD Ryzen AI Max+ 395, and approximately 30 tokens per second on phones. In datacenter settings, a single H100 delivers nearly 15,000 output tokens per second under high concurrency. The model is compatible with tools like llama.cpp, MLX, vLLM, SGLang, and ONNX, and is available on Hugging Face alongside a browser WebGPU demo.

Liquid AI 发布端侧智能体模型 LFM2.5-2.6B Liquid AI 推出面向端侧部署的轻量级模型 LFM2.5-2.6B。该模型体积小于 2.5 GB，支持工具调用与多步工作流，在指令遵循和工具使用任务上表现出可媲美更大模型的性能。 预训练规模约 34T tokens，中间训练阶段将上下文窗口扩展至 128K。 后训练涵盖两轮 SFT、专家教师模型训练、多领域同策略蒸馏及智能体强化学习（Agentic RL）。 在指令遵循和多数工具使用基准中击败或追平高达 9.7B 的对比模型，但在代码能力上落后于大模型。 端侧内存占用低于 2.5 GB，Apple M5 Max 解码速度达 220 tok/s，单张 H100 GPU 高并发输出接近 15K tok/s。 架构与训练流程 LFM2.5-2.6B 经过约 34T tokens 预训练并扩展到 128K 上下文。后训练包含四步：侧重工具使用与网络搜索轨迹的两轮监督微调（SFT）、分领域的专家教师训练、多领域同策略蒸馏（MOPD），以及在实际智能体框架中进行的 Agentic RL。RL 架构将训练引擎、生成动作的 Rollout 引擎与执行沙盒解耦，利用 Harness Proxy 透明捕获 Token 级轨迹。 基准测试表现 在与 gemma-4（5.1B/8B）和 Qwen3.5（4.7B/9.7B）的对比中，LFM2.5-2.6B 在 IFBench、Multi-IF 等指令遵循基准中全部领先，在 ToolSandbox 和 τ³-Bench 等工具使用评测中除 BFCLv4 外均名列前茅；在智能体评测中优于两款 Gemma 模型并与 Qwen 相当。但编程能力（LiveCodeBenchv6）仍明显落后于大尺寸模型。 推理性能与生态支持 该模型支持 llama.cpp、MLX、vLLM、SGLang 与 ONNX。端侧推理方面，Apple M5 Max 达到 220 tok/s，AMD Ryzen AI Max+ 395 为 113 tok/s，手机端可达 30 tok/s；在单张 H100 GPU 高并发场景下，输出速度接近 15,000 tok/s。模型及基础版已上线 Hugging Face，需搭配 transformers>=5.0.0 使用。

LFM2.5-2.6B is built to power capable agents entirely on-device. It supports tool calling and multi-step workflows while staying small and fast enough for everyday hardware, from laptops to phones. This enables developers to deploy agents everywhere, keep data private on the device, and scale usage without a cloud inference bill. Best-in-class agent: Competitive with models 4x larger on tool use, instruction following, and multi-step agentic tasks. Agentic reinforcement learning: Trained inside the most popular agentic harnesses to improve compatibility. Efficient inference : 220 tok/s on an Apple M5 Max and 113 tok/s on an AMD Ryzen CPU, in under 2.5 GB of memory. How we built a reliable agentic model for edge devices LFM2.5-2.6B is pre-trained on ~34T tokens, with a mid-training phase that extends the context window to 128K. Post-training then turns the base model into an agent in four stages: Supervised fine-tuning (SFT): two rounds of SFT, weighted heavily toward agentic data like tool use, web search, and harness trajectories. Teacher specialization: train one specialist teacher per domain (math, code, tool use, and more). Multi-domain on-policy distillation (MOPD): distill the specialist teachers into a single student. Agentic Reinforcement Learning (Agentic RL): run multi-turn RL inside real agent harnesses, where the model learns to work across different tools, system prompts, and multi-turn task environments. The Agentic RL pipeline separates model optimization, inference, and environment execution into distinct components. The Training Engine optimizes the model, while the Rollout Engine generates actions using the latest policy. The RL framework orchestrates the training loop by launching rollouts, collecting trajectories and rewards, and updating the model. Actions are executed within a Sandbox Service , where the Blackbox Harness hosts the agent (e.g., OpenClaw or Hermes Agent) and coordinates interactions with the task environment. The Harness Proxy lets us treat agentic harnesses as black boxes with no modification, while transparently capturing the token-level trajectories needed to reconstruct and validate RL training samples. Benchmark results We evaluated LFM2.5-2.6B against models up to ~4x its size on STEM, instruction following, tool use, and agentic tasks. It is the smallest model in the group, yet it competes with and often beats the rest. Benchmark LFM2.5-2.6B (2.6B) gemma-4-E2B-it (5.1B) gemma-4-E4B-it (8B) Qwen3.5-4B (4.7B) Qwen3.5-9B (9.7B) AA Omniscience -29.50 -74.47 -49.03 -54.30 -50.43 AIME25 51.87 26.33 34.27 49.33 56.07 LiveCodeBenchv6 59.41 54.92 63.77 60.85 69.86 IFBench 59.17 34.08 39.24 48.40 56.47 Multi-IF 80.07 69.44 77.35 55.67 62.55 IFStruct 85.49 64.85 76.65 36.25 78.50 BFCLv4 56.88 36.98 46.39 50.56 60.13 ToolSandbox 77.83 52.40 65.00 75.55 76.44 τ³-Bench Banking 5.67 3.35 4.12 5.45 5.15 Claw-Eval average (EN) 62.85 53.14 58.02 62.28 66.53 PinchBench 68.22 44.24 55.09 71.26 71.45 BrowseComp+ (OpenClaw) 26.89 8.31 15.90 24.46 27.23 For your app, the strengths are instruction following and tool use. LFM2.5-2.6B tops every instruction-following benchmark here, and every tool-use benchmark except BFCLv4, where only the 9.7B Qwen edges ahead. On agentic tasks, it beats both Gemma models and stays even with the Qwens. It also leads on knowledge and stays close on math. Coding is the one place the larger models keep a clear lead, so reach for something bigger there. Inference speed on CPU and GPU LFM2.5-2.6B ships with day-one support across the inference ecosystem, including llama.cpp, MLX, vLLM, SGLang, and ONNX. CPU inference. Due to its efficient LFM2 architecture, LFM2.5-2.6B is the fastest model we tested, with decode speeds of 220 tokens/s on an M5 Max and 113 tokens/s on a Ryzen AI Max+ 395. At 30 tokens/s, it allows you to run capable agents even on a phone. GPU inference. LFM2.5-2.6B is the fastest model in its size class, reaching almost 15K output tokens per second at high concurrency, roughly 1.3B tokens per day on a single H100. How to use LFM2.5-2.6B Reach for LFM2.5-2.6B when you need on-device agents for high-volume workloads. 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 model_id = "LiquidAI/LFM2.5-2.6B" model = AutoModelForCausalLM.from_pretrained( model_id, device_map= "auto" , dtype= "bfloat16" , # attn_implementation="flash_attention_2" # uncomment on a compatible GPU ) tokenizer = AutoTokenizer.from_pretrained(model_id) prompt = "What is C. elegans?" input_ids = tokenizer.apply_chat_template( [{ "role" : "user" , "content" : prompt}], add_generation_prompt= True , return_tensors= "pt" , tokenize= True , ).to(model.device) output = model.generate( input_ids, do_sample= True , temperature= 0.2 , top_k= 80 , repetition_penalty= 1.05 , max_new_tokens= 512 , ) print (tokenizer.decode(output[ 0 ], skip_special_tokens= False )) LFM2.5-2.6B demo Check out this browser demo of LFM2.5-2.6B powering a research agent . The agent helps you research specific questions and generates a summary. Get Started Both LFM2.5-2.6B and LFM2.5-2.6B-Base are available on Hugging Face today. With LFM2.5, we're delivering on our vision of AI that runs anywhere. These models are: Download: LFM2.5-2.6B-Base and LFM2.5-2.6B on Hugging Face. Try: run the WebGPU demo in your browser, no setup needed. Use in your harness: follow our guide on how to run a local agent, like OpenClaw, Hermes Agent, and Pi. We can't wait to see what you build. Citation Please cite this article as: Liquid AI, "LFM2.5-2.6B: Deploy Agents Everywhere", Liquid AI Blog, Aug 2026. Or use the BibTeX citation: @article{liquidAI202626B, author = {Liquid AI}, title = {LFM2.5-2.6B: Deploy Agents Everywhere}, journal = {Liquid AI Blog}, year = {2026}, note = {www.liquid.ai/blog/lfm2-5-2-6b}, }