# Liquid AI releases 24B LFM2.5 model in MLX 6‑bit format on Hugging Face

> The 24‑billion‑parameter LFM2.5 model is now available in a 6‑bit MLX version, runnable with mlx‑lm 0.32.0 or later.

Oossa · 2026-10-13 · https://oossa.com/en/liquid-ai-releases-24b-lfm2-5-model-in-mlx-6-bit-format-on-hugging-face

Liquid AI posted a new version of its LFM2.5‑24B‑A2B model on Hugging Face. The model is converted to MLX format and quantized to 6 bits per weight, keeping the MoE routing gates in full‑precision. It can be loaded with mlx‑lm 0.32.0 or newer; older versions load the model but route experts incorrectly. The model uses a ChatML‑like chat template and supports function calling.

## The facts

- Model size: 24 billion parameters, quantized to 6‑bit weights (group size 64).
- Requires mlx‑lm 0.32.0 or later; earlier versions misroute experts.

## Why it matters

Developers can run a large reasoning model faster on‑device, especially on Apple Silicon machines.

## Sources & references

1. [New model on Hugging Face: LiquidAI/LFM2.5-24B-A2B-MLX-6bit (text-generation)](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-MLX-6bit) – Liquid AI, 2026-10-12
2. [New model on Hugging Face: LiquidAI/LFM2.5-24B-A2B-MLX-5bit (text-generation)](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-MLX-5bit) – Liquid AI, 2026-10-12

Last updated: 2026-10-13