Analysis
Meituan unveils LongCat-2.5 Preview, a 1.6-trillion-parameter agent model with weights still unreleased
Meituan announced LongCat-2.5, a multimodal mixture-of-experts model of about 1.6 trillion parameters. Weights are planned, not published; no benchmarks.
Meituan has announced LongCat-2.5, a multimodal mixture-of-experts model with about 1.6 trillion parameters in total, aimed at long-running work by AI agents. Releases of the weights are described as planned, but we found no repository, date or licence. No benchmarks are published.
The essentials
- What Meituan announces: a mixture-of-experts model, about 1.6 trillion parameters in total and about 48 billion active per token, taking text and images.
- The weights: described as planned to follow by TechNode, citing IT Home; a tracking page says they are not published and that the model is reachable through an API only.
- Not published: benchmarks, licence, release date for the weights, official pricing.
- The date: sources differ, between a preview listed since September 25, 2026 and an announcement dated November 2, 2026.
What does Meituan announce?
According to TechNode (November 2, 2026), which relays IT Home, LongCat-2.5 is a mixture-of-experts model with about 1.6 trillion parameters in total and about 48 billion active per token. By our rounded calculation, that is about 3% of the total (48 / 1,600). zavino.co notes that a "16 trillion" figure seen in some headlines is wrong.
Source: TechNode (November 2, 2026), nocode.mba, AI Weekly.
Tracking pages give a context of about 1 million tokens (ZenMux lists 1.05 million), and a maximum output of 128,000 tokens per nocode.mba, citing Meituan's API documentation. TechNode states neither.
What is it for, according to Meituan?
TechNode says the model targets tasks that need sustained execution by AI agents, especially on complex codebases and full software projects. Training combines extended long-task sequences, large-scale reinforcement learning and failures met by real agents, turned into reproducible training tasks.
The model handles visual and text inputs in a unified token framework. The uses cited are visual coding, creative design and 3D tasks. The outputs are not specified.
Access runs through Meituan's online chat service and its API platform. The tracking page nocode.mba lists a "LongCat-2.5-Preview" on that API, with OpenAI-style and Anthropic-style endpoints. Compatibility with Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code is cited by nocode.mba and ZenMux, not by TechNode.
Where does the release of the weights stand?
TechNode, citing IT Home, writes that publications on GitHub and Hugging Face are planned to follow, with no date or licence. The nocode.mba page says the weights have not been published and the model is reachable through the API only. A web search found no LongCat-2.5 repository. According to several articles, the predecessor LongCat-2.0 was released with open weights on June 30, 2026.
Source: TechNode, nocode.mba, ZenMux, zavino.co.
Which date should we use?
Sources differ, and we do not settle it. nocode.mba and ZenMux list the "LongCat-2.5-Preview" on Meituan's API since September 25, 2026. TechNode dates the announcement of "LongCat-2.5", the multimodal model, to November 2, 2026.
What is not established?
The weights' release date and licence; the context size according to Meituan; official pricing; the multimodal outputs; performance; and which date holds.
Comparison table
| Point | What sources say as of Nov 3, 2026 | Source |
|---|---|---|
| Architecture | Mixture of experts, about 1.6T total, about 48B active per token | TechNode, AI Weekly |
| Inputs | Text and vision, unified token framework | TechNode |
| Context | About 1 million tokens (1.05M at ZenMux); max output 128,000 | nocode.mba, ZenMux |
| Weights | GitHub and Hugging Face planned to follow; not published per a tracking page | TechNode, nocode.mba |
| Benchmarks | None published by Meituan | nocode.mba, zavino.co, AI Weekly |
| Price | ZenMux shows $0.30 per million input tokens and about $1.18 output, shown with a discount; not confirmed by Meituan | ZenMux |
Frequently asked questions
Is LongCat-2.5 open source?
Not to date. Weights are described as planned, but none are published, with no date or licence.
How many parameters does LongCat-2.5 have?
About 1.6 trillion in total, and about 48 billion active per token, per TechNode and AI Weekly.
Does LongCat-2.5 have benchmarks?
None published by Meituan for this preview; claims rest on its own results.
Method and limits
Meituan's posts on X could not be opened. We read the pages through automatic summaries, tested nothing and found no independent benchmark. This is not investment advice.
Sources
- TechNode, "Meituan launches LongCat-2.5 multimodal AI model for extended coding tasks" (November 2, 2026), citing IT Home (verified 2026-11-03)
- AI Weekly (November 2-3, 2026) (verified 2026-11-03)
- nocode.mba, release tracker, secondary source (verified 2026-11-03)
- ZenMux, model page, aggregator (verified 2026-11-03)
- zavino.co, secondary source (verified 2026-11-03)
Spotted an error in this article? Report it · Our corrections policy
