Picture published by the creator on Hugging Face.
LFM2.5-24B-A2B: Run a tool-using reasoning agent on your own device (24B)
Added · Updated
Made by Liquid AI
View profile on Hugging Face (opens in a new tab)Run an AI that thinks step by step and calls tools, on a laptop or your own server.
For developers building private agents that work on device.
Model details
Original model LiquidAI/LFM2.5-24B-A2B
Input Text Output Text
Ways to use it Run locally
Role in this family: Fine-tuned version.
LiquidAI/LFM2.5-24B-A2B-Base (opens in a new tab)What it does
It is a mixture-of-experts model: 24 billion parameters in total, but only about 2 billion are used for each word it produces. It always reasons before it answers and can call tools you describe to it. The creators report 184 tok/s (word pieces per second) on an Apple M5 Max, a context length of 131,072 tokens, and support for languages including English, Arabic, Chinese, Hindi and Spanish.
How to use it
Input → action → output: format your conversation with the model's chat template, optionally with a list of tools written as JSON. The model writes out its reasoning, then either replies in text or returns a tool call for your program to run and send back.
Common use cases
- Give it a list of tools such as web search plus a question, and get back tool calls and a final answer built from the results.
- Run the MLX version on a Mac and chat with a reasoning assistant that runs on your own machine instead of a remote server.
- Fine-tune the separate base model on your own examples to get a version tuned for a specific agent task.
Can I use this?
Requirements Download the weights from Hugging Face; no gate or form is required. Choose the original checkpoint, FP8 for GPU serving, GGUF for llama.cpp, MLX for Apple Silicon, or ONNX. The creators suggest temperature 0.1. License: the custom LFM Open License v1.0, not an OSI open source license; you must include the license when sharing copies and mark changed files.
Caveat Benchmark and speed figures come from the creators. Check the license text before any commercial use.
Worth knowing Text only. The creators do not recommend it for knowledge-heavy tasks on its own; pair it with search or retrieval. Under the LFM Open License v1.0, commercial use is not licensed for companies above a revenue threshold. Licensed under LFM Open License v1.0.
Why it's here
Not trending this week
Last seen on the Hugging Face trending list on 22 October 2026. It stays on BuildTube so you can still find it.
- 603 people have liked it on Hugging Face.
- It was downloaded 38,045 times in the last 30 days.
Numbers checked on 8 November 2026; not refreshed since.
Behind it
View model pageBuildTube has not run or verified this project. Everything above is written from what the creator published.
Made this?
Quick questions
- What is LFM2.5-24B-A2B?
- It is a mixture-of-experts model: 24 billion parameters in total, but only about 2 billion are used for each word it produces.
- What can I use LFM2.5-24B-A2B for?
- Give it a list of tools such as web search plus a question, and get back tool calls and a final answer built from the results; Run the MLX version on a Mac and chat with a reasoning assistant that runs on your own machine instead of a remote server; Fine-tune the separate base model on your own examples to get a version tuned for a specific agent task
- Can I run LFM2.5-24B-A2B on my computer?
- It runs on your own hardware with a model runtime. Choose the original checkpoint, FP8 for GPU serving, GGUF for llama.cpp, MLX for Apple Silicon, or ONNX.
- What do I need to use LFM2.5-24B-A2B?
- Transformers, vLLM or SGLang for the main weights; GGUF, MLX and ONNX versions exist for llama.cpp, Macs and other devices. Exact memory needs are not documented.
- What license does LFM2.5-24B-A2B use?
- Under the LFM Open License v1.0, commercial use is not licensed for companies above a revenue threshold. Licensed under LFM Open License v1.0.
More from this week
- kimi-k3-in-c Run a 2.78-trillion-parameter model on 8 GB of RAM
Get text answers from the huge Kimi K3 AI model using one CPU and no graphics card.
- P-H3-Image-Comfy Make and edit 8-megapixel images in ComfyUI
Turn short text requests into 8-megapixel images and edit pictures on your own graphics card.
- Idol Turn one picture of an object into a 3D Gaussian splat
Idol turns one photo, render or generated image of an object into a 3D Gaussian splat file.
Explore by topic
Post this
Download share imageFor makers
Made this? Show it’s on BuildTube — add the badge to your README or embed the card on your site.
README badge (Markdown)
[](https://buildtube.app/p/hf-liquidai-lfm2-5-24b-a2b/)Embed card (HTML)
<iframe src="https://buildtube.app/embed/hf-liquidai-lfm2-5-24b-a2b/" width="560" height="240" style="border:0;max-width:100%" loading="lazy" title="LFM2.5-24B-A2B on BuildTube"></iframe>