Liquid AI has launched LFM2.5-24B-A2B, a 24-billion-parameter reasoning model with only 2 billion active parameters, designed for fast and private on-device agents. Released today, it is the largest and final model in the LFM2.5 family, which ranges from 230 million to 24 billion parameters. The model is available open-weight on Hugging Face and through Liquid AI's Playground, enabling local fine-tuning and deployment, according to liquid.ai.
The LFM2.5-24B-A2B builds on the previous LFM2-24B-A2B model with enhancements including an expanded vocabulary, longer context window, scaled-up pretraining, and large-scale reinforcement learning. It uses a mixture of experts (MoE) architecture with 40 layers and 64 experts per block, employing top-4 routing and a hybrid of gated short convolutions and GQA. Unlike its predecessor, this model produces an explicit chain of thought before delivering its final answer, improving response quality while maintaining high decoding speed.
Performance benchmarks show LFM2.5-24B-A2B is competitive with Qwen3.6-35B-A3B, a model with about 50% more total and active parameters. It scored 48.09 on LiveClaw and 27.69 on AutomationBench, outperforming Qwen3.6-35B-A3B’s 38.19 and 8.72 respectively. Its architecture allows it to run up to 1.7 times faster than comparable MoEs on-device, achieving 184 tokens per second on an Apple M5 Max chip and up to 21,000 output tokens per second on a single NVIDIA B300 GPU, liquid.ai detailed.
The base and post-trained versions of LFM2.5-24B-A2B are accessible today, with documentation provided to assist developers in running and fine-tuning the model locally. This release marks the completion of the LFM2.5 series, which spans two orders of magnitude in model size and targets efficient, private on-device AI reasoning.