Source context: This brief summarizes a Liquid AI team post hosted on Hugging Face. Performance figures are vendor-reported.

Hugging Face describes LFM2.5-24B-A2B as a 24B-parameter reasoning model with 2B active parameters, aimed at on-device agents that call tools and handle multi-step tasks. That positioning makes local agent deployment its practical focus, rather than universal benchmark leadership. https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b

Hugging Face reports 184 decoding tokens per second on an Apple M5 Max and over 21K output tokens per second on a single B300 at high concurrency; the latter is not a single-user latency measure. Actual application latency and privacy guarantees remain unverified. https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b

Hugging Face acknowledges that larger models retain advantages on knowledge-heavy and coding benchmarks, recommending retrieval or web search for knowledge-heavy workloads. These vendor-reported results suggest a tool-use-focused option, not a demonstrated replacement for stronger coding models; independent reproducibility and workload-specific reliability remain unverified. https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b

Source