Models · Updated 15 Oct, 05:29 am IST

Hugging Face releases LFM2.5-24B-A2B for fast on‑device agents

Why it matters for readers: It’s an accessible example of a large model optimized to behave like a fast, private assistant on regular hardware.

LFM2.5-24B-A2B is a 24B-parameter reasoning model that uses only 2B active parameters per token to lower inference cost.1 2
Hugging Face reports the model is competitive with Qwen3.6-35B-A3B on agentic and tool-use benchmarks, scoring 48.09 on LiveClaw versus Qwen’s 38.19.1 2
The model runs up to 184 tokens/s on an Apple M5 Max and up to 21K output tokens/s on a single NVIDIA B300 GPU.1 2
LFM2.5-24B-A2B expands on LFM2-24B-A2B with a larger context window (up to 131K tokens), doubled vocabulary, more pretraining, and reinforcement learning, and it produces an explicit chain of thought before answers.1 2
Hugging Face made both the base and post-trained open-weight models available on Hugging Face and the Playground, with docs for running and fine-tuning locally.2

Get a brief like this every morning

Uzha reads hundreds of sources and gives you the stories that matter for your work, with every source linked. Free.

13 Oct, 10:35 pm · Hugging Face Blog
Run fast, private agents on-device with LFM2.5-24B-A2B
15 Oct, 05:29 am · Hacker News (100+ points)
LFM2.5-24B-A2B: Built for fast on-device agents