Liquid AI released LFM2.5-24B-A2B. It is a 24B-parameter reasoning model with 2B active parameters. It is built for fast, private on-device agents. It beats Qwen3.6-35B-A3B on LiveClaw. It runs up to 1.7× faster than comparable MoEs on-device. It decodes 184 tokens/s on an Apple M5 Max. It reaches 21K output tokens/s on a single NVIDIA B300 GPU. The base and post-trained models are open-weight and available on Hugging Face. It supports llama.cpp, MLX, vLLM, SGLang, and ONNX for inference. It leads on five of eight agentic and tool-use benchmarks. It is priced at $0.10 per 1M input tokens and $1.00 per 1M output tokens.
What changed
LFM2-24B-A2B had 32K context, 65K vocabulary, and 17T pretraining tokens.
What it unlocks
Running fast, private agents entirely on-device with explicit reasoning chains.
- 24B total parameters
- 2B active parameters
- 131,072 token context window
- 184 tokens/s on Apple M5 Max
What you need to act on it
- a device with Apple M5 Max, AMD Ryzen AI Max+, or NVIDIA B300 GPU
- local deployment via llama.cpp, MLX, vLLM, SGLang, or ONNX
Sources