Meituan's LongCat team introduced LongCat-2.5, a mixture-of-experts model with 1.6T total parameters and about 48B active per token, pitched at long-horizon agentic tasks, visual understanding, and agentic coding — with a chat interface and API live and open weights still to come.
2026-11-02 · by Moe Ameen
On November 2, 2026, Meituan's LongCat team introduced LongCat-2.5, describing it as its latest natively multimodal agentic model. It is a mixture-of-experts model with 1.6 trillion total parameters and roughly 48 billion activated per token, built on the earlier LongCat-2.0 with, per the announcement, stronger native visual understanding and agentic coding — aimed at work ranging from complex coding projects to visual coding.
The multimodal design treats images as discrete tokens rather than bolting a separate vision module onto a text model. Per the announcement, it applies Residual Vector Quantization to native-resolution features from a Semantic-and-Aligned Encoder, so the backbone processes text and images in one token format, and a visual de-tokenizer reconstructs images. The team frames the model around long-horizon agentic tasks — multi-step jobs an agent runs across many turns — and reports scaling an end-to-end agentic reinforcement-learning pipeline across data, environments, and harnesses, with coverage of agentic coding, visual coding, creative design, and 3D tasks.
On availability, a chat interface and an API are live; the GitHub and Hugging Face buttons are labelled "coming soon," so open weights are not out yet. Reporting puts the limited-time API price at roughly $0.30 per million input tokens and $1.20 per million output tokens, with the API documentation listing a 1M-token context window and up to 128K output tokens. Meituan says the model was trained on domestically produced AI accelerators and uses W4A4 (4-bit weight and activation) inference, and reports it is compatible with agentic coding harnesses including Claude Code, Codex, and Hermes.
Specs, prices, and benchmark figures on a day-old release move fast and often come through secondary coverage, so treat the numbers above as a snapshot and confirm the current figures on LongCat's official channels before planning around them.
Be honest about the fit: LongCat-2.5 will not make your TikToks. It is an agentic coding and multimodal reasoning model — useful for building things and running agents, not for generating the finished, platform-native content a creator posts. The news underneath the news is the one that keeps repeating: the raw model layer gets cheaper and stronger every few weeks, and none of that closes the gap between an idea and a week of on-brand posts shipped across every feed. That gap is the job [Kompozy](/) exists for, and it is the thing no frontier model release changes.
So here is the useful move today. If you are technical, wire a cheap agentic model like this into a research or planning step — have it scan a transcript, draft an outline, or reason over source material. Then hand the actual production to the engine built for it: Kompozy turns one source into [Persona Shorts](/glossary/persona-shorts) and avatar video, [Clipped Shorts](/glossary/clipped-short) reframed per platform with word-synced captions, brand-exact [Carousel Posts](/glossary/hyperframes) and quote graphics, plus a blog article and an email newsletter — all held to one voice by the [Persona Brief](/glossary/persona-brief), then scheduled and published across the eight social platforms plus blog and email on [Autopilot](/glossary/autopilot) behind a per-post review gate. Kompozy runs its generation on Claude and OpenAI under the hood, not LongCat; the point is that a new model is a new ingredient, and the finished meal still needs a kitchen. For the full picture of what the model actually does, see our [LongCat-2.5 tool page](/ai-tools/longcat-2-5).
LongCat-2.5 is a natively multimodal mixture-of-experts AI model from Meituan's LongCat team, introduced on November 2, 2026. It has 1.6 trillion total parameters with roughly 48 billion activated per token, builds on LongCat-2.0, and is aimed at long-horizon agentic tasks, agentic coding, and visual understanding. A chat interface and API are available; open weights are listed as coming soon.
Not directly. LongCat-2.5 is built for agentic coding and multimodal reasoning — writing code, building interfaces from images, running multi-step agent tasks — not for generating finished short-form video, images, or carousels for a feed. A creator would use it as a reasoning or building tool and a separate engine like Kompozy to produce and publish the actual content.
Not yet, as of the November 2026 launch. The announcement lists GitHub and Hugging Face releases as 'coming soon,' so for now LongCat-2.5 is reachable through its chat interface and a hosted API rather than a self-hostable model. Confirm the current status on LongCat's official channels, since this is likely to change.
Secondary coverage of the launch puts the limited-time API price at roughly $0.30 per million input tokens and $1.20 per million output tokens, with cached input reported far cheaper, a 1M-token context window, and up to 128K output tokens. API-platform pricing is authoritative and can change, so verify the current rates before budgeting.