Meituan moves LongCat-2.5 out of preview, a 1.6T-parameter multimodal model for long agent tasks

Meituan's LongCat team has released the full version of LongCat-2.5, a multimodal model built for AI agents that work on long tasks such as entire software projects. The launch post is dated November 1 and was widely reported on November 2. The model is not entirely new: a LongCat-2.5-Preview went live on Meituan's API platform on September 25 with the same headline size. What changes now is that the preview label is gone and Meituan has published its own benchmark comparisons.

What Meituan is shipping

According to TechNode, LongCat-2.5 has about 1.6 trillion total parameters, of which roughly 48 billion are activated per token, and it handles images and text within a single token framework. Meituan pitches it for visual coding, creative design and 3D work as well as conventional programming. The company says it trained the model on longer task sequences with large-scale reinforcement learning, and turned failures observed in real agents into reproducible training tasks.

The model is available in Meituan's chat service and through its API. The API documentation now lists LongCat-2.5 as its only model, with a 1M-token context window, up to 128K output tokens, and both an OpenAI-style and an Anthropic-style endpoint, which lets it plug into existing coding agents. Pricing and any free allowance for the final version could not be confirmed.

Benchmarks and hardware claims

FourWeekMBA, which read the launch post, reports that Meituan's table puts LongCat-2.5 close to GPT-5.6-Sol and Claude Opus 5 on several agent tests, such as 88.4 on Terminal Bench v2.1 against 89.1 for Claude Opus 5, but well behind on others, at 38.5 on Terminal Bench v4.0 against 53.9. Meituan describes some rival figures as indicative only, and no independent evaluation has been published.

The post also says Meituan optimized long-context training and 4-bit inference on domestic AI accelerators, without naming the chip vendor. That continues the line of LongCat-2.0, which the company released as open source at the end of June, saying it had been trained on domestic chips.

What is still missing

Unlike LongCat-2.0, the new model has no public weights yet: GitHub and Hugging Face releases are marked as coming soon, with no date and no stated license. Until those arrive, LongCat-2.5 is effectively an API product, and its performance claims rest on Meituan's own testing.