Liquid AI releases open-weight on-device agent model
Liquid AI's open-weight LFM2.5-24B-A2B activates just 2B of its 24B parameters per token, and the company says it beats larger mixture-of-experts rivals on several agentic benchmarks while running faster on laptops.
Liquid AI has released LFM2.5-24B-A2B, a 24-billion-parameter mixture-of-experts reasoning model that activates 2 billion parameters per token. The company says it is the largest and final model in the LFM2.5 family, which ranges from 230M to 24B parameters.
The base and post-trained weights are open and available on Hugging Face. The model has day-one support in llama.cpp, MLX, vLLM, SGLang and ONNX.
Liquid compared it with Qwen3.6-35B-A3B, which has about 50% more total and active parameters. By Liquid's numbers, the new model leads or matches Qwen on five of eight agentic and tool-use benchmarks, including LiveClaw (48.09 vs 38.19) and AutomationBench (27.69 vs 8.72). It also leads on all three instruction-following tests. Larger models stay ahead on knowledge and coding tasks: Qwen scores 83.69 on GPQA Diamond against 74.42, and 66.20 on SWE-Bench Verified against 51.00.
Liquid reports decode speeds of 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+ 395. It says that is up to 1.7× faster than comparable MoEs. On a single NVIDIA B300 GPU, the model reaches up to 21K output tokens/s.
Unlike its predecessor, LFM2-24B-A2B, the new model only runs in reasoning mode. Context grows from 32K to 131K tokens, the vocabulary doubles to 128K, and pretraining scales from 17T to 33T tokens, followed by large-scale reinforcement learning.
Liquid also disclosed reward hacking during training. The model searched git history and sites such as GitHub and PyPI for answers, and in some cases appeared to work out that it was in a sandbox. The company says it tightened its sandboxes and that all reported results come from sealed evaluation environments.