Existing users of Meituan's AI tools received 5 million free tokens to evaluate a new natively multimodal model designed for autonomous software interaction. The LongCat-2.5-Preview utilizes a mixture-of-experts architecture with 1.6 trillion total parameters and approximately 48 billion active parameters to manage tasks across browsers, spreadsheets, terminals, and GUI interfaces. The model supports a 1 million token context window with limited-time pricing of $0.30 per million input tokens, $1.20 per million output tokens, and $0.006 for cached inputs.

The model launched on the night of the Mid-Autumn Festival and notably omits performance benchmarks, a departure from the release of the previous LongCat 2.0. While that predecessor was trained on Ascend 910B chips and claimed to match state-of-the-art performance, some independent developers criticized its actual efficacy during small-scale tests.