BriefsNewsHugging FaceAvailable now

Liquid AI releases LFM2.5-24B-A2B, an open-weight on-device reasoning model

Liquid AI's new 24B-parameter model uses 2B active parameters per token and is built for tool-calling agents that run on your own hardware.

Announced by Hugging FacePublished Checked

Summary

Liquid AI released LFM2.5-24B-A2B, a reasoning model that activates only 2B of its 24B parameters per token. It is aimed at agents that call tools and run locally. The weights are open and on Hugging Face, with support for llama.cpp, MLX, vLLM, SGLang and ONNX.

What changed

  • New 24B-parameter reasoning model with 2B active parameters, built for on-device agents
  • Context window grows from 32K to 131K tokens compared with the previous version
  • Reported speeds: 184 tokens/s on an Apple M5 Max and 114 tokens/s on a Ryzen AI Max+ 395
  • Leads or matches rivals on five of eight tool-use benchmarks; larger models stay ahead on knowledge and coding
  • Open weights for both the model and the Base version, available on Hugging Face

Availability

Available now

“Both LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base are open-weight and available on Hugging Face today:”

Founder takeaway

If you build agents that call tools and want them private and running on your own machine, this is worth a test through a local OpenAI-compatible server. For knowledge-heavy or coding work, the page itself suggests adding retrieval or web search.

Sources

  1. Hugging Face — Run fast, private agents on-device with LFM2.5-24B-A2BPublished · Checked

Researched and drafted with AI from the official sources above, then checked automatically before publishing. Spotted a mistake? Tell us.

Recent

More briefs

All briefs

  1. Published Tool

    Available now

    Google Research releases Deep Review, an agentic AI tool for checking documents

    Deep Review, an AI system that checks long technical documents, is now public, with two free reviews per week and an API for heavier use.

    Google Research · Announced

  2. Published News

    Available now

    GitHub's AI classifier for secret scanning cuts false alerts and scan time

    GitHub describes a fine-tuned ModernBERT model that sorts real credentials from lookalikes, replacing an LLM-based pipeline in secret scanning.

    GitHub · Announced

  3. Published Workflow

    Available now

    AWS adds SpaceXAI's Grok 4.7 AI model to Amazon Bedrock, with agent tests

    Grok 4.7 is on Amazon Bedrock, and AWS tested how its reasoning effort setting affects accuracy and token use in AI agents.

    AWS · Announced

One email, every Monday

Briefs are published here every weekday. Monday Heist is the weekly email: one useful AI workflow, the files to use it, and the catch. Free, and you can leave in one click.

One email. Every Monday.