Liquid AI released LFM2.5-24B-A2B on Tuesday, introducing a 24-billion-parameter reasoning model designed to run local agents using two billion active parameters per token. Hugging Face published the technical disclosure, confirming that both the base model and an instruction-tuned variant are available under open weights.
Engineers structured the architecture with 40 layers, 64 experts per block, and top-4 routing, combining gated short convolutions with grouped query attention. The design expands the context window from 32,000 to 131,000 tokens while doubling the vocabulary from 65,000 to 128,000 tokens. Training expanded the pre-training corpus from approximately 17 trillion to 33 trillion tokens, followed by a mid-training phase that added 2 trillion tokens at 32,000 context and 300 billion tokens at 131,000 context. Post-training added supervised fine-tuning, two stages of multi-domain on-policy distillation, and agentic reinforcement learning inside harnesses including OpenClaw, Hermes Agent, Pi, and OpenCode.
Benchmark scores show the model competitive with larger MoEs possessing roughly 50 percent more total and active parameters. On the LiveClaw evaluation, LFM2.5-24B-A2B scored 48.09, outperforming Qwen3.6-35B-A3B at 38.19, Gemma-4-26B-A4B-it at 41.25, and Nemotron-3.5-Lightning at 37.72. The system recorded 27.69 on AutomationBench, exceeding the next-best competitor score of 8.72. Larger models retained advantages on knowledge-intensive evaluations including GPQA Diamond and coding evaluations like SWE-Bench Verified, where Qwen scored 66.20 compared to Liquid AI’s 51.00.
Hardware tests showed decoding speeds of 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395. A single Nvidia B300 GPU produced more than 21,000 output tokens per second under high concurrency. Deployment relies on compatibility with Transformers version 5.0.0 or later, alongside runtime support across vLLM, SGLang, MLX, ONNX, and llama.cpp.
