Meituan releases LongCat 2.5 for visual, long-running agents
The full model reaches Meituan’s API and Vercel AI Gateway. Its long-context capacity and hardware-efficiency results describe different parts of the agent workload.
Editorial visualizationMeituan’s LongCat 2.5 is now available for developers building agents that work with images, code and multi-step engineering tasks. The official API changelog dates the release to November 2, 2026, following LongCat 2.5 Preview on September 25. This is the full model’s service launch, rather than its first public preview.
Vercel added LongCat 2.5 to AI Gateway on November 2, giving developers another route to the model under the name meituan/longcat-2.5. Vercel lists text and image inputs and describes support for coding and multi-step tool use. Its availability announcement confirms a distribution channel; the claimed improvements in visual programming and longer workflows are attributed to Meituan.
The API parameter reference specifies a context limit of 1,048,576 tokens, counting input and output together. That gives an agent room to carry code, documents and task history within a request. It is a capacity limit, not a measurement of how accurately the model uses every part of a long history. The same reference says temperature is fixed at 1.0 with thinking enabled and 0.7 with thinking disabled, overriding other requested values.
The hardware work behind longer tasks
Meituan’s technical release describes training on domestic AI accelerators with less than 80 GB of memory each. It keeps key exchanges inside high-bandwidth supernodes and shards model state to reduce memory pressure. In tested long-sequence supervised fine-tuning, its optimized configuration delivered 86% higher per-device throughput than a distributed-data-parallel baseline.
For serving the model, Meituan reports four-bit expert weights and activations—the W4A4 scheme—using native INT4 operations. Its online stress tests compare this with an eight-bit W8A8 INT8 baseline. Prefill processes the input; decode generates the answer token by token.
| Serving measure | Reported change |
|---|---|
| Prefill throughput | +43.3% |
| Decode throughput | +42.77% |
| Time to first token | −15.91% |
| Time per output token | −19.72% |
These are configuration-specific efficiency results, not a speed comparison with a rival model. Meituan reports comparable quality after quantization on evaluated downstream tasks; that does not establish equivalence on every workload.
For teams evaluating the release, the useful test combines both sides: whether a supervised agent completes a real engineering task correctly, and what latency and resource use that task requires. The service launch makes that assessment possible; the developer’s infrastructure measurements do not answer it on their own.
Publication history
— First published.
No published corrections or earlier versions.
How this was made
UpThat uses AI-assisted research and writing. Source checking and publication follow our editorial review workflow. See the original documents above to assess the reporting.
How we use AI ↗
