# Liquid AI releases LFM2.5-24B-A2B, a 24B on-device reasoning MoE with 2B active

> LFM2.5-24B-A2B is a 24B-parameter reasoning model with about 2.45B active, built for agents on-device, under the LFM1.0 licence.

Published 2026-10-12

Canonical: https://shipped.anth.us/open-weights/liquidai-lfm2-5-24b-a2b/

Liquid AI published LFM2.5-24B-A2B, which the card calls the largest and final model in the LFM2.5 family. It is a text-only reasoning model with 24.0B total and 2.45B active parameters, 40 layers (30 double-gated short convolution blocks and 10 grouped-query attention blocks) and 64 experts per MoE block with top-4 routing. Context is 131,072 tokens. It covers 16 languages, listed on the card.

The card says it was trained on 36 trillion tokens and that it always thinks before it answers. The maker claims 184 tokens per second on an Apple M5 Max and up to 21,000 output tokens per second on one NVIDIA B300, and it reports a score of 48.09 on LiveClaw against 38.19 for Qwen3.6-35B-A3B. These are Liquid AI's own figures.

The licence is listed as lfm1.0, a custom licence. A base model and an FP8 checkpoint sit alongside it.

The card gives no release date; the date used is the Hugging Face repo creation date.

A small active size does the heavy lifting here. Check the licence before shipping it.

## Sources
- [Model card on Hugging Face](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B)
- [Hugging Face repo metadata (creation date, licence, size)](https://huggingface.co/api/models/LiquidAI/LFM2.5-24B-A2B)

## Editor's first pass (AI-assisted)
- Real 4: Ungated weights plus FP8 and GGUF builds; the LFM1.0 licence is custom.
- Useful 3.5: Cheap reasoning tokens suit local agents; the card quotes speeds on Apple and NVIDIA hardware.
- Notable 3.5: 550 likes; the card calls it the largest and final LFM2.5 model.
- Novel 3.5: A hybrid of short-convolution blocks and attention rather than a plain transformer.