Liquid AI releases LFM2.5-2.6B on-device agent model
Liquid AI has released LFM2.5-2.6B, a 2.6B-parameter model built to run agent workflows on phones and CPUs, and reports benchmark results competitive with Gemma 4 and Qwen3.5 models up to nearly four times its size.
Liquid AI announced LFM2.5-2.6B, a small model trained for agentic work such as planning, tool calling and multi-step tasks, and designed to run entirely on-device. The base model (LFM2.5-2.6B-Base) and the post-trained model are both available on Hugging Face. The company argues that running agents locally removes per-token costs, which would let developers run many agents in parallel and keep background tasks going without paying for each token.
The model was pre-trained on about 34T tokens. Liquid doubled the vocabulary to 128K by extending the existing tokenizer rather than retraining from scratch, to improve support for non-Latin scripts. Mid-training included a 128K-context extension phase.
Post-training had four stages:
- Supervised fine-tuning (SFT): two rounds, with a data mix about seven times larger than the one used for LFM2.5-8B-A1B.
- Specialist teachers: separate expert models trained with RLVR for areas including math, code, tool use and long context.
- Multi-domain on-policy distillation: the specialists' skills were distilled back into a single model.
- Agentic RL: GRPO training inside real agent harnesses, including Hermes Agent and OpenClaw, using verl, SGLang and FSDP.
In Liquid's own evaluations, LFM2.5-2.6B scored 77.83 on ToolSandbox, 56.88 on BFCLv4 and 51.87 on AIME25. It beat Qwen3.5-4B on most of the published benchmarks but trailed Qwen3.5-9B on AIME25, LiveCodeBench v6 and BFCLv4. No independent results have been reported yet. The blog post does not give latency or memory figures for specific devices.