Meituan's natively multimodal agentic model — a 1.6T-parameter mixture-of-experts system built for long-horizon agentic tasks, agentic coding, and visual understanding.
Last verified · 2026-11-02 · by Moe Ameen
LongCat-2.5 is a natively multimodal mixture-of-experts model from Meituan's LongCat team, introduced on November 2, 2026. It carries 1.6 trillion total parameters with roughly 48 billion activated per token, builds on the earlier LongCat-2.0, and is aimed squarely at long-horizon agentic tasks — multi-step jobs an agent runs across many turns — alongside agentic coding and visual understanding. For the scored verdict on the model itself, see the [LongCat-2.5 review](/reviews/longcat-2-5).
What makes it "native" multimodal is how it handles images. Per the announcement, it treats images as discrete tokens using Residual Vector Quantization on native-resolution features from a Semantic-and-Aligned Encoder, so the model processes text and images in one token format rather than attaching a separate vision encoder, with a visual de-tokenizer reconstructing images. That design is what underpins its visual-coding claims — turning a screenshot or design into working front-end code — and it reports strength on web-from-image and text-to-3D style tasks.
On access, a chat interface and an API are live; GitHub and Hugging Face releases are listed as "coming soon," so open weights are not out yet. Secondary coverage puts the limited-time API price near $0.30 per million input tokens and $1.20 per million output, with a 1M-token context window and up to 128K output tokens. Meituan says it was trained on domestically produced AI accelerators and uses W4A4 inference, and reports compatibility with agentic coding harnesses including Claude Code, Codex, and Hermes.
The honest scope line, because it decides whether this tool is for you: LongCat-2.5 is a reasoning, coding, and agent model — not a generator of finished social video, images, or carousels. It can help you build and automate; it does not render the content you post. Confirm current specs and pricing on LongCat's official channels, since a day-old release moves quickly.
LongCat-2.5 is a builder's model, so the cleanest way to pair it with [Kompozy](/) is ship-then-market. Use LongCat-2.5 to actually build the thing — turn a mockup into a working web app with its visual coding, run an agent that refactors a codebase, generate a 3D asset or an internal automation. That is net-new product work the model is genuinely good at. What it cannot do is produce the content that gets anyone to notice what you shipped: the launch reel, the demo carousel, the founder-voice posts, the blog walkthrough, the newsletter. That is the exact handoff to Kompozy.
Point Kompozy at your launch notes, a demo recording, or the product's own copy and it generates the whole marketing set: [Persona Shorts](/glossary/persona-shorts) and avatar video fronting the launch, [Clipped Shorts](/glossary/clipped-short) cut from a demo and reframed to 9:16, 1:1, and 16:9 with word-synced captions, brand-exact [Carousel Posts](/glossary/hyperframes) walking through features, quote graphics, photo posts, plus a blog article and an email newsletter — every piece governed by the [Persona Brief](/glossary/persona-brief) so a solo builder sounds like a brand, not a changelog. Then it schedules and publishes across the eight social platforms plus blog and email on [Autopilot](/glossary/autopilot) behind a per-post review gate. For reference, Kompozy's own generation runs on Claude and OpenAI rather than LongCat; the division of labor is the point — LongCat builds the product, Kompozy builds the audience for it. Compare how other frontier models slot into the same chain on the [GPT-6](/ai-tools/gpt-6) and [Claude Opus 5.5](/ai-tools/claude-opus-5-5) pages.
Long-horizon agentic tasks and coding. LongCat-2.5 is Meituan's natively multimodal mixture-of-experts model (1.6T parameters, ~48B active) built for multi-step agent workflows, agentic and visual coding — including turning screenshots or designs into front-end code — and visual understanding. It is a reasoning and building model, not a generator of finished social-media video or images.
No. It is not a video or image generator for content. Its multimodal strength is understanding images and producing code, 3D, and reasoning outputs, not rendering reels, carousels, or avatar videos for a feed. To make and publish that content you would use a dedicated engine like Kompozy, which generates 18 content formats and publishes across nine destinations.
Through its chat interface and a hosted API as of the November 2026 launch; open weights on GitHub and Hugging Face are listed as 'coming soon.' Reported limited-time API pricing is roughly $0.30 per million input tokens and $1.20 per million output, with a 1M-token context window. It is reported compatible with Claude Code, Codex, and Hermes. Verify current access and pricing on LongCat's official channels.
Use LongCat-2.5 to build the product — code from a design, agent automations, 3D assets. Then use Kompozy to market it: point Kompozy at a demo or your launch notes and it generates avatar video, clipped shorts, carousels, quote graphics, a blog, and a newsletter in one brand voice, and publishes them across the eight social platforms plus blog and email. One builds; the other distributes.