Liquid AI Releases 2.6B Parameter Edge-Side Model, 3 Benchmarks Exceed Qwen3.5-9B

According to TrendForce Beating Monitor, Liquid AI has released LFM2.5-2.6B and made the model weights open-source. This on-device Agent model has only 2.6 billion parameters, with a quantized memory footprint of about 2.5GB, capable of achieving a generation speed of approximately 30 Tokens per second even on mobile devices.

In official benchmarks, it outperforms Qwen3.5-9B in three tests: tool invocation, multi-turn instruction following, and structured output. The latter, with 9.7 billion parameters, is nearly four times larger than Liquid AI's model. However, Qwen3.5-9B still maintains an edge in code-related and some complex Agent tasks.