Liquid AI released LFM2.5-24B-A2B, an open-weight reasoning model built for agents that run on the device rather than in a data center.
The company published base and post-trained checkpoints on Hugging Face and called the model the largest and final entry in its LFM2.5 family, which now runs from a 230M parameter version up to this one.
Two changes separate it from its predecessor. The context window grew fourfold, and an in-place tokenizer expansion doubled the vocabulary for better handling of non-Latin scripts.
The architecture keeps the earlier design: a mixture-of-experts stack that activates a small fraction of its parameters per token. Total size is 24B, but only about 2.45B are active at any moment, which is what makes local inference plausible.
The training budget was 36 trillion tokens, and the model covers sixteen languages. Unlike the version it replaces, it is reasoning-only and does not emit plain conversational output.
Crucially, the weights are public. That lets enterprises run agents on hardware they control, avoiding per-token cloud costs and keeping sensitive data on premises.
The tradeoff is capability. Serving a smaller active-parameter budget locally means accepting less raw reasoning power than a frontier API offers.
