LFM2.5-8B-A1B

About

LFM2.5-8B-A1B is Liquid AI's reasoning-tuned, general-purpose Mixture of Experts (MoE) model designed for on-device deployment - an on-device personal assistant that chains tool calls and follows complex instructions on all devices. It is a sparse hybrid of 18 double-gated convolution blocks and 6 Grouped-Query Attention (GQA) full-attention layers (24 layers total): 8.3B total parameters with only 1.5B active per token (top-4 routing across 32 experts), with Rotary Position Embedding (RoPE) base frequency 5M and a 128,000-token context window.

It was trained with a 38-trillion-token pre-training budget and large-scale reinforcement learning, and its vocabulary was expanded in place to 128K entries. LFM2.5-8B-A1B is text-only, speaks English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese and Spanish, uses a ChatML-like chat template, and supports Pythonic function calling in a four-step tool-use flow. It is "compressed performance": competitive with much larger dense and MoE models on instruction following and agentic tasks, and is the fastest model in its size class - up to 18.5K output tokens per second at high concurrency on a single H100 - with day-one support for llama.cpp, MLX, vLLM and SGLang. A 328M speculative decoding drafter (LFM2.5-8B-A1B-DSpark) pairs with it for ~2.5x faster decoding with identical outputs.

Recommended for agentic workflows, tool use, structured outputs, multilingual assistants and on-device personal-assistant applications; not the best fit for heavy programming or knowledge-intensive question answering without retrieval. Released May 28, 2026 under the LFM Open License v1.0.

Training Data 38 trillion tokens pre-training, post-training with SFT and RL