According to CoinMeta, Liquid AI has released a model with 2.6 billion parameters named LFM2.5-2.6B, and has made the model weights available for public access. This model, with only 2.6 billion parameters, occupies about 2.5GB of memory after quantization and can still achieve a generation speed of approximately 30 token per second on mobile devices. Official evaluations show that this model outperforms Qwen3.5-9B, which has 9.7 billion parameters (almost four times as many), in three tests: tool invocation, multi-round instruction following, and structured output. However, Qwen3.5-9B still leads in terms of code quality and in some complex agent tasks.