NewsLiquid AI2 min read
LFM2.5-24B-A2B: An Open Model for Agents on Your Laptop
Liquid AI releases LFM2.5-24B-A2B, a 24-billion-parameter open model that activates only 2 billion per token, built for private agents on your own computer.
Key points
- On 13 October Liquid AI released LFM2.5-24B-A2B, an open-weight model you can download from Hugging Face.
- It has 24 billion parameters but uses only 2 billion at each step, which is why it is fast on a laptop.
- Liquid AI says it generates 184 tokens per second on an Apple M5 Max.
- It is a reasoning model: it always thinks before answering, with up to 131,000 tokens of context.
⚠ Pending verification. We have not yet confirmed these details with an official source. Check them before relying on them:
- How much memory it needs: Liquid AI gives no figures, only that it uses the least of the models compared.
- The exact terms of the LFM 1.0 licence, such as any limits for large companies: the model card only names it.
Most powerful models live in the cloud: you type, your text travels to a server and the answer comes back. LFM2.5-24B-A2B, which Liquid AI released on 13 October, is built for the opposite: letting an agent work on your own computer, fast and without your data leaving it.
The trick in the name: 24B, but A2B
The name sums up the idea. The model has 24 billion parameters (24B), but at each step it activates only 2 billion (A2B). It is a mixture of experts: inside there are 64 specialised “experts”, and for each token only 4 are picked. It knows as much as a large model, but each answer costs the compute of a small one.
What Liquid AI says about performance
| Figure | According to Liquid AI |
|---|---|
| Speed on an Apple M5 Max | 184 tokens per second |
| Speed on an AMD Ryzen AI Max+ 395 | 114 tokens per second |
| Context | 131,000 tokens (up from 32,000) |
| LiveClaw (agent tasks) | 48.1, vs 38.2 for Qwen3.6-35B-A3B |
| SWE-Bench Verified (coding) | 51.0, below larger models (66.2) |
Liquid AI admits that larger models beat it on coding and hard science questions. Its strengths are agent tasks and following instructions. All the figures are its own.
How to try it
It is an open-weight model: you download it from Hugging Face and run it with the usual local tools, such as llama.cpp, MLX (on Mac) or vLLM. There are quantised versions that need less memory. It is released under Liquid AI’s own “LFM 1.0” licence; read it before using the model in a product.
It is a reasoning model: it always writes out its chain of thought before answering, so replies take a little longer to start.
What it means for you
- If privacy matters to you, a model like this lets you build an assistant or agent that sends nothing to the internet.
- If you have a recent Mac or a powerful laptop, you can try it without paying for an API. If not, a cloud model is still more practical.
- If you compare open models, read the table with care: Liquid AI chooses who it compares against. Our AI model guide explains how to read these comparisons.
