LiquidAI says LFM2.5-24B-A2B has 24 billion total parameters and 2 billion active per token, and describes it as a reasoning model that calls tools and handles multi-step tasks. Its context window is 131,072 tokens. These are the observed facts behind the release, not an independent assessment of model quality.
The mechanism is: The model’s sparse activation and harness-compatible serving aim to make local reasoning and tool calls practical; the blog does not isolate how much each component contributes. LiquidAI describes staged post-training that includes agentic reinforcement learning in harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode, and lists support for llama.cpp, MLX, vLLM, SGLang, and ONNX. The company reports 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, while also acknowledging larger models lead on some knowledge and coding tests. The blog does not isolate how much each component contributes.
Why now: the weights are available on Hugging Face today, so developers can test the local-agent claim rather than wait for a hosted access path. The first implication is that teams can benchmark a lower-active-parameter option against their own local agent workload before committing to a serving stack. The first watch condition is whether independent, reproducible tests confirm useful tool-use quality and speed across hardware and harnesses. LiquidAI’s benchmark and throughput figures remain vendor-reported; they do not establish performance on every workload, device, or configuration. This starts a distinct analytical record because no prior New Runtime record was supplied; it should not be read as a continuation of another record.