LFM2.5-24B-A2B, a 24 billion parameter model for AI agents that run on devices, is released (liquid.ai)

Writing language: KoreanRead in the original language

- This model is a reasoning model that uses only 2 billion of its 24 billion total parameters at a time to answer. It is built to handle tasks by calling tools on personal devices such as laptops, and its weights can be downloaded from Hugging Face. - Liquid AI said in its announcement that, compared with Qwen3.6-35B-A3B, this model is competitive with or ahead on 5 of 8 tool use and agent benchmarks. In the same table, its SWE-Bench Verified coding score is 51.00, lower than Qwen3.6-35B-A3B's 66.20. - The claims of a processing speed of 184 tokens per second on Apple M5 Max and up to 1.7 times faster come from the developer, based on measurements Liquid AI made on its own benchmark machine Pipette. The source contains no independent verification.

Found on

Hacker News ↗ / 37 votes / 7 comments (checked 2026-10-15 09:56 KST)

Have you run into something like this? Tell us in a line or two what you did. You do not need to know the product.

1 comment

Editorial opinionstartupxo 1h ago

O: Developers who want to run a local agent on a laptop could download the model from Hugging Face and try a short task, such as a morning briefing, themselves by following the local agent guide in the original article. X: The benchmark rests on generation conditions set by Liquid AI and measurements taken on its own equipment, so it is a claim from the announcing party. It is also worth noting that its SWE-Bench score for coding tasks is lower than that of Qwen3.6-35B-A3B.

Writing language: Korean

Related reads

1 comment