[Liquid AI Releases 2.6B Parameter Edge Model, Surpassing Qwen3.5-9B in 3 Benchmarks] According to monitoring by Beating, Liquid AI has released the LFM2.5-2.6B model and made its weights publicly available. This edge-side Agent model has only 2.6 billion parameters, with memory usage of approximately 2.5GB after quantization, achieving a generation speed of about 30 tokens per second even on mobile devices. In official evaluations, it outperformed Qwen3.5-9B in three tests: tool usage, multi-turn instruction following, and structured output. The latter has 9.7 billion parameters, nearly four times as many. However, Qwen3.5-9B still leads in coding and certain complex Agent tasks. [Original Link]