The reasoning model has 24 billion total and 2 billion active parameters per token. It extends context to 131,072 tokens and targets tool calling and agent loops on local hardware.
Base and post-trained weights are on Hugging Face, with support announced for llama.cpp, MLX, vLLM, SGLang and ONNX. A Liquid API is also available. It is an explicit-reasoning model, not an interchangeable non-reasoning mode of its predecessor.
Liquid reports gains on selected agent benchmarks but also shows larger models retaining advantages on knowledge and coding tasks. Two billion active parameters do not mean only two billion weights must be stored. Open weights are not unrestricted licensing; review the LFM license.
Local models offer targeted trade-offs in speed, autonomy and cost rather than universal superiority.