LFM2.5-2.6B release notes / what's new
Liquid AI has announced the release of LFM2.5-2.6B, a model specifically engineered to power capable agents entirely on-device. By combining a small footprint with high performance in tool calling and multi-step workflows, LFM2.5-2.6B enables developers to deploy agents on hardware ranging from laptops to smartphones, ensuring data privacy and eliminating cloud inference costs.
Technical Architecture and Training Pipeline
LFM2.5-2.6B is pre-trained on approximately 34 trillion tokens, with a mid-training phase that extends its context window to 128K. The transition from a base model to a specialized agent is achieved through a four-stage post-training process:
- Supervised Fine-Tuning (SFT): Two rounds of SFT focusing heavily on agentic data, including tool use, web search, and harness trajectories.
- Teacher Specialization: The creation of specialist teacher models for specific domains such as math, code, and tool use.
- Multi-domain On-Policy Distillation (MOPD): Distilling the knowledge from these specialist teachers into a single student model.
- Agentic Reinforcement Learning (Agentic RL): Multi-turn RL conducted within real agent harnesses to improve the model's ability to operate across various tools, system prompts, and task environments.
The Agentic RL pipeline utilizes a Training Engine for model optimization, a Rollout Engine for action generation, and an RL framework to orchestrate the loop. Actions are executed in a Sandbox Service using a Blackbox Harness (such as OpenClaw or Hermes Agent) and a Harness Proxy to capture token-level trajectories for RL training validation.
Performance Benchmarks
LFM2.5-2.6B competes with models up to four times its size, often outperforming them in instruction following and tool use.
Key Benchmark Results
| 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 |
| 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 |
| Claw-Eval average (EN) | 62.85 | 53.14 | 58.02 | 62.28 | 66.53 |
| BrowseComp+ (OpenClaw) | 26.89 | 8.31 | 15.90 | 24.46 | 27.23 |
LFM2.5-2.6B leads in every instruction-following benchmark listed and nearly every tool-use benchmark, with only the 9.7B Qwen model edging it out on BFCLv4. While it remains competitive on agentic tasks and knowledge, larger models maintain a clear lead in coding tasks.
Inference Speed and Hardware Compatibility
LFM2.5-2.6B is designed for high-efficiency inference across various hardware and software ecosystems, including llama.cpp, MLX, vLLM, SGLang, and ONNX.
- CPU Inference: The model achieves decode speeds of 220 tokens/s on an Apple M5 Max and 113 tokens/s on a Ryzen AI Max+ 395, requiring under 2.5 GB of memory.
- GPU Inference: At high concurrency, the model reaches nearly 15K output tokens per second, which equates to roughly 1.3B tokens per day on a single H100 GPU.
Deployment and Usage
LFM2.5-2.6B is available on Hugging Face in both base and tuned versions. It is compatible with transformers>=5.0.0. Developers can implement the model using the AutoModelForCausalLM and AutoTokenizer classes from the Hugging Face Transformers library.