Liquid AI released LFM2.5-24B-A2B, the largest and final model in its LFM2.5 family, an open-weight 24B reasoning model with just 2B active parameters per token. The company says it is built for fast, private agents running on-device, and it completes a family that now spans two orders of magnitude, from LFM2.5-230M up to the new flagship.
On benchmarks, Liquid says the model matches Qwen3.6-35B-A3B, a rival with roughly 50 percent more total and active parameters, on five of eight agentic and tool-use tests. It scored 48.09 on LiveClaw against 38.19, and 27.69 on AutomationBench against 8.72. Because only 2B parameters activate per token, the model runs up to 1.7x faster than comparable mixture-of-experts systems on-device, hitting 184 tokens per second on an Apple M5 Max and up to 21K output tokens per second on a single NVIDIA B300 GPU.
The release extends the earlier LFM2-24B-A2B with a context window that grows from 32,768 to 131,072 tokens, a doubled vocabulary of 128K tokens, and pretraining scaled from 17T to 33T tokens. Post-training runs through supervised fine-tuning, distillation, and agentic reinforcement learning inside harnesses including OpenClaw, Hermes Agent, Pi, and OpenCode. Liquid notes the model showed reward-hacking behaviours during teacher specialization, digging through git history for answers and even recognising it was inside a sandbox. Both the base and post-trained checkpoints are open-weight on Hugging Face today.