Meituan’s new preview targets long-running work across coding tools, browsers and visual interfaces. The launch specifications are substantial; the evidence for reliable end-to-end execution is still to come.
Meituan launched LongCat-2.5-Preview on September 25, 2026, with 1.6 trillion total parameters, approximately 48 billion active parameters and a one-million-token context window. The new capability documented for this release is image understanding. Meituan positions the model for extended tasks across terminals, browsers, desktop interfaces, spreadsheets and design tools. It is available through LongCat Chat and the LongCat API platform. Sources: launch announcement, official changelog, API quick start.
The scale is familiar; the visual input is the update
The 1.6T total and roughly 48B active figures describe a mixture-of-experts model: only a fraction of its parameters participate in the computation for a given token. That fraction is about 3% by parameter count. It does not mean the model fits in the memory footprint of a 48B-parameter dense model; the full expert set still has to be stored and served. For another 1M-context sparse model, see our Step 5 Preview architecture analysis. Meituan gave the same scale figures for LongCat-2.0, which also offered a one-million-token context window. The 2.5 launch therefore should not be read as a new jump in advertised parameter count or context length. Sources: 2.5 announcement, LongCat-2.0 introduction.
The more consequential change is the new image understanding capability described in Meituan’s September 25 changelog. The company says the preview can interpret images for visual questions, summaries and reasoning. That matters for agents because a screenshot, a chart or a design mockup can become part of the working context, rather than requiring a separate vision model to describe it first.
There is an important boundary to the claim. Meituan names terminals, browsers, GUIs, spreadsheets and design tools as target environments, but the model itself does not become a computer-control system merely by seeing an image. An agent runtime must still supply tools, permissions, state tracking and a way to execute and verify actions. Our comparison of coding-agent harnesses explains why the same model can behave differently across tool environments. Meituan’s integration guides show how to connect the model to developer tools; they do not establish a measured success rate for complete workflows.
What the API actually exposes
The model ID is LongCat-2.5-Preview. Meituan documents both an OpenAI-compatible chat completion endpoint and an Anthropic-compatible Messages endpoint. Its quick-start guide lists a 1M-token context window and 128K maximum output for both 2.5 and 2.0. The chat API also exposes a thinking setting that can be enabled or disabled. Sources: API quick start, chat completions reference.
The currently published chat-completions reference describes message content as a plain-text string and illustrates only a text request. It does not show an image payload in that reference. Developers planning screenshot-heavy workflows should check the platform’s current image-input format with a small call before migrating a production pipeline. This is a documentation gap, not evidence that the announced image capability is absent. Source: chat completions reference.
The limited-time price makes long trials practical
Meituan’s official pricing page lists the following US-dollar rates per million tokens:
| Token type | Standard rate | Limited-time rate |
|---|---|---|
| Uncached input | $0.75 | $0.30 |
| Cached input | $0.015 | $0.006 |
| Output | $2.95 | $1.20 |
At the promotional rate, one million uncached input tokens would cost $0.30 before output charges. That is a pricing calculation, not a promise that a million-token request will be fast, accurate or available without rate limits. Cached-token savings also depend on the workload achieving cache hits; our Claude Fable 5.1 cost analysis shows why the full task bill can differ from the advertised cache rate. Meituan says the platform’s displayed prices and billing records are authoritative, so the limited-time rates should be checked when running a trial. Sources: official pricing, API limits.
In the launch announcement, Meituan also offered five million free tokens to existing users, with balances visible on the usage page. The company says existing token packages work with both 2.5 and 2.0; its FAQ confirms support for both models. The announcement does not establish that every new account receives the same grant.
The unanswered question is task completion
A large context window can keep more files, logs and screenshots in one request. It does not, by itself, show that the model can retrieve the right detail after hundreds of steps, notice when a tool action failed or recover without corrupting a spreadsheet. Likewise, image understanding is an input capability; reliable GUI work requires precise action selection and verification.
Meituan’s launch post and September 25 changelog do not provide a 2.5-specific benchmark table or a reproducible, multi-application agent evaluation. As our SWE-bench audit discusses, a coding score alone does not settle production reliability. The right test is a complete task with a known end state: give the agent a repository issue plus a browser-based reference and a spreadsheet output, then record completion rate, human interventions, elapsed time, token use and errors across repeated runs. Compare those results with LongCat-2.0 under the same tools and limits. Until such evidence is available, claims of stronger coding and long-horizon performance remain the vendor’s positioning. Sources: launch announcement, official changelog.
LongCat-2.5-Preview is immediately interesting for teams that already run agents through OpenAI- or Anthropic-style APIs: they can trial the new visual capability without redesigning their entire client. The verdict on whether it can finish long tasks more reliably than 2.0 will come from those end-to-end results, not from the trillion-parameter headline.
Frequently asked questions
Is LongCat-2.5-Preview open source?
Meituan’s launch materials reviewed here describe chat and API access. They do not announce public 2.5 model weights. LongCat-2.0 was released separately as an open model; that status should not be assumed for 2.5.
Does the one-million-token window mean one million tokens of output?
No. The API guide lists a one-million-token context window and a separate 128K maximum output. The usable input budget in a request depends on how much output space is reserved.
Who gets the five-million-token grant?
The announcement says existing users receive five million free tokens. Users can check their actual balance on the platform usage page.
