Liquid AI Releases LFM2.5-24B-A2B MoE Model for Local Agents

Liquid AI has released LFM2.5-24B-A2B, an MoE reasoning model with 2.45B active parameters out of a total of 24B. This model aims to run local agents on laptops and desktops, supporting a 131K token context window. It is designed to mandatorily generate a reasoning process, including the <think> tag, before producing any answer.

Performance and Benchmarks

According to Liquid AI's own measurements, the model demonstrated superiority over Qwen3.6-35B-A3B in tool-calling and instruction-following performance. It scored 27.69 on AutomationBench (Qwen: 8.72) and 96.75 on IFEval. However, it scored lower than the Qwen model in knowledge (GPQA Diamond: 74.42) and coding (SWE-Bench Verified: 51.00).

Execution Speed and Resources

Based on llama.cpp, the Q4_K_M quantized model achieved a decoding speed of 184 tok/s on an Apple M5 Max, with memory usage of approximately 15GB. On an NVIDIA B300 GPU, it showed a throughput of 21,070 tok/s, approximately 1.67x faster than Qwen when processing 1,024 concurrent requests. Korean tokenizer efficiency was improved by reducing the number of tokens by 16.7% compared to previous models through vocabulary expansion.

License and Usage Restrictions

The model is licensed under the LFM Open License v1.0, which restricts commercial use by companies with annual revenue exceeding $10 million. Commercial use is permitted for research and companies with annual revenue below $10 million. It can be run in environments with Transformers 5.0.0 or higher and can integrate with local agent harnesses such as Hermes Agent.