Liquid AI released LFM2.5-24B-A2B on October 13, opening the weights of a reasoning model designed for agents running on local hardware. Both base and post-trained versions are available.
What “2B active” means
The model has 24 billion parameters, with about two billion active for each token. That distinction matters: sparse activation reduces computation, but does not make the full model equivalent to a two-billion-parameter download. Deployment still requires memory for its weights and runtime.
The update also extends the context window to 131,072 tokens and expands the vocabulary. Liquid AI reports improved results on several tool-use evaluations, alongside weaknesses relative to larger models on some knowledge and coding tests. These are company-run comparisons, rather than independent findings.
The deployment question
For developers, the practical test is whether a local agent completes useful work within the available memory, latency and power budget.
Running inference locally can give teams greater control over model processing. Privacy still depends on the complete workflow, including any external services an agent calls.
Source: Liquid AI