Run fast, private agents on-device with LFM2.5-24B-A2B
huggingface ·
LiquidAI has released LFM2.5-24B-A2B, a 24-billion-parameter reasoning model designed for fast and private on-device agent execution. Despite its large size, the model utilizes a Mixture of Experts architecture that activates only 2 billion parameters per token, significantly reducing inference costs. The update features an expanded context window of 131K tokens and a doubled vocabulary to improve non-Latin language handling. Training involved scaling pre-training data to approximately 33 trillion tokens and incorporating multi-stage post-training focused on agentic tasks. Benchmarks show the model outperforming larger competitors like Qwen3.6-35B-A3B in tool use and instruction following, particularly scoring 48.09 on LiveClaw. Inference speeds are highly efficient, reaching 184 tokens per second on Apple M5 Max hardware and over 21K tokens per second on NVIDIA B300 GPUs. The model is open-weight and compatible with major inference ecosystems including llama.cpp, MLX, and vLLM. Users can integrate it into local agent harnesses such as Hermes Agent or OpenClaw for reliable edge deployment.
huggingface