Hugging Face releases LFM2.5-24B-A2B, a 24B-parameter reasoning agent with 2B active parameters for on-device use

LiquidAI’s LFM2.5-24B-A2B is a 24-billion-parameter reasoning model designed for fast, private agents running on-device. It activates only 2 billion parameters per token, reducing compute cost and enabling continuous operation on consumer hardware.

The model is competitive with Qwen3.6-35B-A3B—a model with roughly 50% more total and active parameters—on agentic and tool-use tasks, achieving a score of 48.09 on LiveClaw compared to 38.19 for Qwen3.6-35B-A3B.

LFM2.5-24B-A2B uses a mixture-of-experts architecture with 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid of gated short convolutions and grouped-query attention. Its context window has been expanded from 32K to 131K tokens, and its tokenizer vocabulary increased from 65K to 128K tokens for improved non-Latin tokenization.

Pre-training scaled 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, multilingual text, and long-context data. Post-training included supervised fine-tuning, teacher specialization, multi-domain on-policy distillation (both non-agentic and agentic), and agentic reinforcement learning within harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode.

Across 16 benchmarks covering tool use, agentic workflows, knowledge, instruction following, math, and coding, LFM2.5-24B-A2B is the smallest model in its comparison group yet remains competitive. It leads or matches competitors on five of eight tool-use and agentic benchmarks—including AutomationBench, where it scores more than three times the next-best model—and on all three instruction-following benchmarks.

Inference speed tests show the model decoding 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395, up to 1.7 times faster than the next-best model while using the least memory. On GPU, it reaches over 21,000 output tokens per second at high concurrency on a single NVIDIA B300.

The model ships with day-one support across the inference ecosystem, including llama.cpp, MLX, vLLM, SGLang, and ONNX. It is available under the identifier LiquidAI/LFM2.5-24B-A2B on Hugging Face and is compatible with transformers versions 5.0.0 and above.