Meituan says it has released LongCat-2.5, a natively multimodal agentic model with 1.6 trillion total parameters and about 48 billion active. The company aims it at long-horizon agentic work, from complex coding projects to visual coding such as creative design and 3D tasks.

Image shared by @Meituan_LongCat on X
Image shared by @Meituan_LongCat on X

Image credit: @Meituan_LongCat on X

What changed from the preview

Meituan first introduced LongCat-2.5-Preview in a September 25 post, listing the same 1.6T and roughly 48B active parameters, a 1M-token context window, native multimodality, and API and chat access. Superpowerdaily describes the November 2 release as moving the coding-agent work from that preview into a hosted release on the chat service and developer API, per Superpower Daily.

Eneralabs listed preview pricing at $0.30 per million input tokens and $1.20 per million output tokens, according to Eneralabs. Inside AI reports that Meituan has not announced API pricing for this release.

Access and training

Inside AI reports the release came out on Monday, November 2, 2026, and is available now through Meituan's online chat service and API platform. It also says GitHub and Hugging Face releases of the weights are planned, with no date given.

The launch page meta description says the 1.6-trillion-parameter model was trained entirely on domestic Chinese chips. Meituan says it expanded training on long task sequences using large-scale reinforcement learning, and that it turns failures seen in real agents into reproducible training tasks.

Benchmarks

Reports differ. Inside AI says Meituan has not disclosed benchmark results. FourWeekMBA reports a 31-benchmark table comparing LongCat-2.5 with GPT-5.5, GPT-5.6-Sol, Claude Opus 4.8, Claude Opus 5, GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and says Meituan ran its Terminal Bench tests inside Claude Code's harness. FourWeekMBA's own count puts LongCat-2.5 ahead of GPT-5.6-Sol on 20 of 27 shared tests. Superpowerdaily says Meituan has not published project-completion scores.

Our read: treat that win count as one outlet's tally of a single table, and wait for independent evaluations before comparing the model with closed rivals.

Sources