LFM2.5-24B-A2B
Liquid AI released LFM2.5-24B-A2B, an open-weight 24B mixture-of-experts reasoning model with about 2B active parameters, built for tool-calling agents on local hardware.
It keeps the LFM2-24B-A2B architecture (40 layers, 64 experts per block with top-4 routing). It grows the context window from 32K to 131K tokens and doubles the vocabulary from 65K to 128K tokens. Pre-training grew from about 17T to about 33T tokens, and the model is now reasoning-only. Post-training ends with agentic reinforcement learning inside harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.
- Date
- Tuesday 13 October 2026
- Lab
- Liquid AI
- Kind
- open-weights
- Access
- open weights
- Price
- $0.05 in / $0.70 out per M tokens on OpenRouter
Figures
| Measure | Value | Measured by |
|---|---|---|
| LiveClaw | 48.09 versus 38.19 for Qwen3.6-35B-A3B | company |
| AutomationBench | 27.69 versus 8.72 for Qwen3.6-35B-A3B | company |
| BFCLv4 | 65.01 versus 65.00 for Qwen3.6-35B-A3B | company |
| SWE-Bench Verified (OpenCode) | 51.00 versus 66.20 for Qwen3.6-35B-A3B | company |
| GPQA Diamond | 74.42 versus 83.69 for Qwen3.6-35B-A3B | company |
| Decode speed on Apple M5 Max | 184 tokens/s 114 tokens/s on Ryzen AI Max+ 395 | company |
| Output throughput on a single NVIDIA B300 | over 21K tokens/s at high concurrency | company |
| Parameters | 24B total, 2.45B active per the OpenRouter listing | OpenRouter |
Weights are under the LFM 1.0 License, and the Hugging Face tags list the license as other. Benchmarks are Liquid's own, and the model trails Qwen3.6-35B-A3B on knowledge and coding tests.
Sources
- huggingface.co/blog/LiquidAI/lfm25-24b-a2b
- openrouter.ai/liquid/lfm-2.5-24b-a2b
- huggingface.co/LiquidAI/LFM2.5-24B-A2B-FP8
This record was partly confirmed: some claims could not be checked on 13 October 2026. How we check