Liquid AI 2.6B Model Outperforms Qwen3.5-9B in Three Benchmarks

Liquid AI releases open-source LFM2.5-2.6B, beating Qwen3.5-9B in tool use and structured output despite 4x smaller size, signaling efficiency gains.

Woofun AI reports that Liquid AI has released the LFM2.5-2.6B model with open-source weights. This on-device Agent model contains 2.6 billion parameters and occupies approximately 2.5GB of quantized memory, achieving a generation speed of roughly 30 Tokens per second on mobile devices.

Official benchmarks indicate the model surpasses Qwen3.5-9B in tool invocation, multi-turn instruction following, and structured output. Despite being nearly four times smaller than the 9.7 billion parameter Qwen3.5-9B, it demonstrates superior performance in these specific areas, although Qwen3.5-9B retains advantages in code-related tasks and certain complex Agent operations.