Liquid AI has released LFM2.5-24B-A2B, a Mixture-of-Experts reasoning model with 24 billion total parameters. The "A2B" designation means only 2 billion parameters are active for each token, a design that reduces compute cost while keeping a large knowledge base.
The model is aimed at on-device agents. According to the announcement, it scored 48.09 on LiveClaw, a benchmark for agentic tasks, and reached 184 tokens per second decoding on an Apple M5 Max. That speed suggests the model can run locally with responsive performance.
Because the model is open-weight and requires only 2B active parameters, it could make sophisticated reasoning more practical on personal hardware rather than relying on cloud APIs. This is a single vendor announcement, so independent benchmarks or comparisons with other on-device models are not yet available.