# Liquid AI releases LFM2.5-24B-A2B reasoning model on Hugging Face

> The 24‑billion‑parameter model, with only 2 billion active parameters, is tuned for fast on‑device agents and tool use.

Oossa · 2026-10-13 · https://oossa.com/en/liquid-ai-releases-lfm2-5-24b-a2b-reasoning-model-on-hugging-face

Liquid AI added LFM2.5-24B-A2B to Hugging Face on Oct 12 2026. It is a 24 B‑parameter MoE model that activates just 2 B parameters per token, making inference cheap and fast. The company says it matches Qwen3.6‑35B‑A3B on agentic and tool‑use benchmarks while running up to 1.7× faster on‑device, reaching 184 tok/s on an Apple M5 Max and over 21 K tok/s on a single NVIDIA B300 GPU.

## The facts

- 24 B total parameters, 2.45 B active (≈2 B per token)
- Scores 48.09 on LiveClaw, beating Qwen3.6‑35B‑A3B’s 38.19

## Why it matters

Developers can run sophisticated reasoning agents locally without cloud costs, enabling faster, private AI tools.

## Sources & references

1. [New model on Hugging Face: LiquidAI/LFM2.5-24B-A2B (text-generation)](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B) – Liquid AI, 2026-10-12
2. [Liquid AI Releases LFM2.5-24B-A2B: A 2B-Active MoE for On-Device Agents](https://www.marktechpost.com/2026/10/14/liquid-ai-lfm2-5-24b-a2b/) – MarkTechPost, 2026-10-14

Last updated: 2026-10-14