Liquid AI Releases LFM2.5-24B-A2B for On-Device Agents

Liquid AI unveiled LFM2.5-24B-A2B on October 13, 2026, describing it as the final model in its LFM2.5 family and its largest flagship to date. The open-weight model is aimed at on-device AI agents, according to aitimes.com.

The model carries 24 billion total parameters and activates about 2 billion per token through a mixture-of-experts architecture. Liquid AI claims up to 1.7x faster on-device inference than comparable MoE models. It generated 184 tokens per second on Apple M5 Max and 114 tokens per second on AMD Ryzen AI Max+ 395; on a single Nvidia B300 GPU it handled over 21,000 output tokens per second under high concurrency.

On agent and tool-use benchmarks, the model scored higher than or similar to Qwen3.6-35B-A3B on 5 of 8 tests. It reached 27.69 on AutomationBench against 8.72 for the competitor, and 48.09 on LiveClaw against 38.19. Qwen3.6-35B-A3B still led on SWE-bench Verified, a benchmark from background testing, with 66.20 versus 51.00.

Pretraining data grew from 17 trillion to 33 trillion tokens, and context length expanded from 32,768 to 131,072 tokens. Vocabulary size doubled from 65,000 to 128,000 to improve non-Latin text handling. The model builds on the earlier LFM2-24B-A2B and follows Liquid AI's releases ranging from 230M to 2.6B parameters.

LFM2.5-24B-A2B is available on Hugging Face and through llama.cpp, MLX, vLLM, SGLang and ONNX. API pricing is 0.10 USD per million input tokens, 1 USD per million output tokens and 0.05 USD per million cached input tokens.