AI BriefingLFM2.5-2.6B Released
Key point
Liquid AI released a 2.6B on-device agent model.
Details
Liquid AI released LFM2.5-2.6B. It is an on-device agent model designed to run on general devices such as laptops and smartphones while supporting tool calling and multi-step workflows.
The model was pre-trained on approximately 34 trillion tokens, with the context length extended to 128K through mid-training. It was developed through agent-specific SFT, domain-specific teacher models, multi-domain on-policy distillation, and Agentic RL based on real agent harnesses.
- 220 tok/s on Apple M5 Max
- 113 tok/s on AMD Ryzen CPU
- Memory usage under 2.5GB
- Integration with real agent harnesses such as OpenClaw and Hermes Agent
In the released benchmarks, the 2.6B model was compared against larger Gemma and Qwen series models in STEM, instruction following, tool use, and agentic task performance.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.