Run fast, private agents on-device with LFM2.5-24B-A2B
Hugging Face introduces LFM2.5-24B-A2B, a 24-billion-parameter reasoning model optimized for on-device agent deployment with only 2 billion active parameters per token, enabling cost-effective, private, and fast local execution.
Hugging Face released LFM2.5-24B-A2B, a 24-billion-parameter reasoning model designed to operate with just 2 billion active parameters per token, allowing continuous agent execution on local hardware. The model generates an explicit chain of thought before producing final answers, enabling multi-step tasks and tool calls within standard agent frameworks. Its architecture retains the 40-layer structure with 64 experts per MoE block and top-4 routing, while expanding the context window from 32K to 131K tokens and doubling the vocabulary to 128K tokens for improved non-Latin tokenization.
Training for LFM2.5-24B-A2B scaled pre-training from approximately 17 trillion to 33 trillion tokens, followed by mid-training with 2 trillion tokens at 32K context and 300 billion tokens at 131K context, emphasizing reasoning, math, code, instruction following, and multilingual data. Post-training transformed the base model into an agent through five stages, with tool calls and multi-turn interactions integrated throughout. Evaluations against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning across 16 benchmarks showed LFM2.5-24B-A2B as the smallest model yet competitive with larger MoEs, particularly excelling in tool use and instruction following.
LFM2.5-24B-A2B leads or matches competitors on five of eight tool-use and agentic benchmarks, including AutomationBench where it scores more than three times higher than the next-best model, and outperforms all three instruction-following benchmarks. While larger models retain advantages on knowledge-heavy and coding tasks, LFM2.5-24B-A2B is positioned as a strong foundation for reasoning, tool use, and reliable on-device agent operation, especially when paired with retrieval or web search for knowledge-intensive workloads.
The model launches with day-one support across multiple inference backends, including llama.cpp, MLX, vLLM, SGLang, and ONNX. On-device performance reaches 184 tokens per second on an M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, up to 1.7 times faster than the next-best model while using the least memory. On GPU, it achieves over 21,000 output tokens per second at high concurrency on a single B300, making it the fastest in its size class.