Liquid AI releases LFM2.5-24B-A2B, a 24B reasoning model with just 2B active parameters

Liquid AI published LFM2.5-24B-A2B on its Hugging Face blog on October 13, a reasoning model with 24 billion total parameters but only 2 billion active per token, built for fast, private agents on-device. The company says it thinks before it answers, calls tools, and works through multi-step tasks inside popular agent harnesses.

Because only 2B parameters are active per token, each reasoning token is cheap, so teams can run agents continuously on their own hardware instead of sending every request to the cloud. That keeps sensitive data local while containing inference cost.

On agentic and tool-use tasks, Liquid AI says the model is competitive with Qwen3.6-35B-A3B, which has roughly 50% more total and active parameters, scoring 48.09 on LiveClaw against 38.19 for the rival. For speed, the company reports 184 tokens per second on an Apple M5 Max and up to 21,000 output tokens per second on a single NVIDIA B300 GPU.

Architecturally it keeps the LFM2-24B-A2B design: 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid of gated short convolutions and grouped-query attention. The main changes are a context window that grows from 32K to 131K tokens and an in-place tokenizer expansion that doubles the vocabulary from 65K to 128K for better non-Latin tokenization.

Unlike its predecessor, LFM2.5-24B-A2B is a reasoning-only model that produces an explicit chain of thought before its final answer. Pre-training scaled from about 17 trillion to 33 trillion tokens, with mid-training adding 2 trillion tokens at 32K context weighted toward reasoning, math, code, instruction following and multilingual text, plus 300 billion tokens at 131K context.

Post-training turns the base model into an agent in five stages: supervised fine-tuning, teacher specialization, two rounds of multi-domain on-policy distillation (one non-agentic, one agentic), and agentic reinforcement learning inside real harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.

The notable signal is that Liquid AI trains its agentic reinforcement learning inside the same harnesses customers use, blurring the line between model and tooling. Weights are open on Hugging Face, so teams weighing on-device privacy against cloud convenience now have a concrete model to test on long-context, multi-step tool use.