Liquid AI releases a 24B agent model with 2B active

Watch

Lab claim

On October 13, 2026, Liquid AI released LFM2.5-24B-A2B, a 24 billion parameter model with 2 billion active and a 131,000 token context. The card's license name is LFM 1.0.

How to read this claim

Measurement — No independent rerun of LiveClaw or the token rates is cited. LFM 1.0 is not Apache 2.0.
Bill — List input and output per million tokens on the vendor page — a loop pays both again.
Door — No independent rerun of LiveClaw or the token rates is cited. LFM 1.0 is not Apache 2.0.

Keep the habit · field card with prices and worked checks

What changed

The previous on-device item in this briefing embeds files. LFM2.5 is a reasoning model that Liquid says calls tools, with 2 billion parameters active.

What it introduces

Open weights for a small-active-parameter agent model. The license on the card is LFM 1.0.

Who benefits Anyone who wants a tool-using model on a machine they already have, and who can read a license that is not Apache 2.0.

Try this

Point a local server at the weights and run one tool-use task you already score. Keep Liquid's 184 token figure as Liquid's until you measure your own machine. The license name on the card is LFM 1.0.

What we do not know

No independent rerun of LiveClaw or the token rates is cited. LFM 1.0 is not Apache 2.0.

Evidence

Primary Liquid AI. Reported by Liquid AI. The speed and benchmark figures are Liquid's.

Independent The LiveClaw score and the token rates are Liquid's. No independent rerun is cited.

Sources

Liquid AI Primary · October 13, 2026
Weights

Still unclear? · Start here · ranks, prices, unknown No independent rerun of LiveClaw or the token rates is cited. LFM 1.0 is not Apache 2.0.

The record

What happened

On October 13, 2026, Liquid AI released LFM2.5-24B-A2B, a reasoning model with 24 billion parameters and 2 billion active. The post says the context window is 131,000 tokens, up from 32,000 on the previous model. Liquid measures 184 tokens per second on an Apple M5 Max and up to 21,000 output tokens per second on one NVIDIA B300. On LiveClaw, Liquid reports 48.09 for this model and 38.19 for Qwen3.6-35B-A3B. The Hugging Face card names the license LFM 1.0.

Why it matters

EmbeddingGemma 2 searches on the device. This model is meant to take tool-using steps on the device, with 2 billion parameters active so the reasoning tokens stay cheap. The LiveClaw gap and the token rates are Liquid's measurements. The license is not Apache, so the card has to be read before the weights go into a product.

Liquid says post-training included agentic reinforcement learning inside harnesses it names as OpenClaw, Hermes Agent, Pi, and OpenCode. The LiveClaw and token-rate figures are Liquid's.