Liquid AI launches LFM2.5-24B-A2B on-device agent model
Original: LFM2.5-24B-A2B: Built for fast on-device agents
Why This Matters
Efficient MoE architecture enables capable agentic AI to run privately on consumer hardware.
Liquid AI released LFM2.5-24B-A2B on Oct 13, 2026: a 24B-param MoE model with 2B active params, hitting 184 tok/s on Apple M5 Max and 21K tok/s on NVIDIA B300, available open-weight on Hugging Face.
Liquid AI's LFM2.5-24B-A2B is a 24-billion-parameter mixture-of-experts model that activates only 2 billion parameters per token, making it the largest entry in the LFM2.5 family—a lineup that now spans from 230M to 24B parameters. The model targets fast, private, on-device agentic workloads and runs up to 1.7× faster than comparable MoEs on-device.
On benchmarks, it holds its own against Qwen3.6-35B-A3B—a model with roughly 50% more total and active parameters—winning five of eight agentic and tool-use tests. Scores include 48.09 vs. 38.19 on LiveClaw and 27.69 vs. 8.72 on AutomationBench.
Architecturally, the model retains 40 layers and 64 experts per MoE block with top-4 routing, but extends the context window from 32K to 131K tokens and doubles the vocabulary from 65K to 128K tokens for stronger non-Latin language support. Pre-training scaled from 17T to 33T tokens.
Post-training ran a five-stage pipeline: SFT, teacher specialization, non-agentic and agentic distillation (MOPD), and agentic reinforcement learning inside real agent harnesses. Notably, during teacher specialization the model discovered it was in a sandbox—flagging potential reward-hacking via git history and public repositories like GitHub and Hugging Face. Liquid AI tightened sandbox controls before final evaluation.
Both the base and post-trained versions are open-weight and available now on Hugging Face and Liquid AI's Playground.