LFM2.5-8B-A1B
Liquid AI releases LFM2.5-8B-A1B, an on-device mixture-of-experts model with 8B total parameters and 1B active, trained on 38T tokens with RL post-training. Optimized for fast tool calling and instruction following on consumer hardware with a 128K context window.
Specifications
| Lab | Liquid AI |
|---|---|
| Context window | 128K tokens |
| Input price | $0.00/1M |
| Output price | $0.00/1M |
| Release | May 2026 |
Strengths
- Extremely efficient on-device inference with only 1B active parameters
- Strong tool calling and instruction following (IFEval 91.84, BFCLv3 64.36)
- Doubled vocabulary (128K) with significant multilingual tokenization improvements
- Day-one support for llama.cpp, MLX, vLLM, and SGLang
- Competitive with much larger dense and MoE models on agentic tasks
Weaknesses
- Limited reasoning depth (AIME25: 42.53) compared to frontier models
- Small active parameter count limits complex multi-step reasoning
- Not yet available on major API providers like OpenRouter
Best for
- On-device AI assistants and personal agents
- Tool-calling workflows on consumer hardware
- Multilingual applications with non-Latin scripts
- Edge deployment where latency and efficiency matter most
In Depth: LFM2.5-8B-A1B
Summary
LFM2.5-8B-A1B is an AI model from Liquid AI.
Released in May 2026. It currently appears in the Overall category on LMRank and 1 other category. It takes text input and produces text output, with a context window of 128K tokens. Input pricing is $0.00 per 1M tokens and output is $0.00 per 1M tokens on OpenRouter.