Meituan released the weights for LongCat-2.5 on November 9, 2026, one week after introducing the model. The Chinese food delivery and e-commerce company published the checkpoint on Hugging Face under the MIT license, giving developers a new way to run a model that can inspect screenshots and video, work through a codebase, and write or revise software over multiple steps. At 1.6 trillion total parameters, it is one of the largest open-weight models ever released — and the permissive license means teams can use it commercially, modify it and redistribute it without asking anyone.
The timing is no accident. This week has been crowded with agent announcements, from Google’s Gemini 4 Carbon for long agent work to the always-on AI coworker Databricks is building with Genie. LongCat-2.5 is the open-weight answer to that same question: what happens when the agent doing the work is one you can download, inspect and run yourself?
What LongCat-2.5 Is, and What Shipped on November 9
LongCat-2.5 is the successor to LongCat-2.0, Meituan’s earlier flagship from July 2026, and it inherits both its size and its license. It is a multimodal mixture-of-experts model: 1.6 trillion total parameters, with about 48 billion active for each token, so only a subset of the model’s experts works on any given request. It represents visual information as tokens and processes it alongside text, which is what makes it an agent model rather than a chat model — it can look at something, reason about it and act on it.
What shipped on November 9 is the full checkpoint, not just API access. The weights sit in a Hugging Face repository of 208 files in BF16 format, taking up about 3.56 terabytes. Meituan has also published a technical write-up and an API guide, and the model can be tried without downloading anything at longcat.ai. One wrinkle worth flagging: the Hugging Face repository metadata displays 1.8 trillion parameters, while Meituan’s technical description says 1.6 trillion. Teams sizing a deployment should work from the files and the exact serving format, not the headline number.
There is also an important distinction in the language around this release. The weights are open, but the training data and the full training pipeline are not published, and the only code Meituan points to is the code for serving the model. Open-weight is the accurate description, not open-source — though for most engineering teams, downloadable weights under MIT are the part that matters.
An Agent That Can See the Screen
The strongest case for testing LongCat-2.5 is a job that requires both visual judgment and coding. Think of a developer handing an agent a screenshot of a broken checkout flow, asking it to find the relevant components in a repository, and having it propose a patch and check the result in a browser. The API accepts screenshots, charts, images and video, with one constraint worth noting: images can be embedded in a request or supplied by URL, while video must be supplied by URL, which complicates workflows that use internal screen recordings.
Meituan’s own demos point in the same direction. The company highlighted a car-launch site, a fantasy game interface and a shop management site built with the model, alongside an interactive fluid simulation and a 3D exhibition. For frontend and product engineering, the sensible first test is a visual regression with a known fix: give the agent the failing screenshot, access to a test repository and a way to render the patched page, then check whether it finds the cause, makes a targeted change and preserves keyboard navigation and responsive behavior. A design-to-code task is a second good test — but the output should be scored on whether it follows the team’s component library and interaction rules, not just whether it looks right.
The intended workload is long, multi-step agent work: large codebase development, full-project builds, visual design and 3D tasks, with the model reinforced through long training on agentic tasks where the AI works for hours without a human guiding it step by step.
The Benchmarks: Big Gains Over LongCat-2.0, Mixed Against Frontier Rivals
The clearest comparison in Meituan’s evaluation is with its predecessor. LongCat-2.5 scores 88.4 versus 70.8 on Terminal Bench v2.1, 64.6 versus 9.7 on DeepSWE v1.1, and 58.3 versus 33.1 on the SWE Atlas codebase question-answering test. The jump on DeepSWE, from 9.7 to 64.6, is the headline number — and the one to treat with the most care, since scores without an asterisk in Meituan’s model card are the company’s own in-house measurements, which the card calls indicative only. There is no independent evaluation yet.
The rival comparisons are more mixed. LongCat-2.5 comes within one point of GPT-5.6 Sol and Claude Opus 5 on Terminal Bench v2.1, and it leads its comparison group on 3DCodeBench with a score of 55.4. But on the newer Terminal Bench v4.0 it scores 38.5, against 53.9 for Claude Opus 5, 58.2 for GPT-6.1 Sol and 64.8 for Claude Opus 5.5. DeepSeek-V4.1-Flash leads the group on DeepSWE with 74.2, and Claude Opus 5 leads on Vision2Web with 80.2 against LongCat’s 78.6.
There is a telling detail in Meituan’s choice of reference points. The company compared LongCat-2.5 against GPT-5.6 Sol and Claude Opus 5, but their direct successors were already available when the model was introduced: OpenAI released GPT-6.1 Sol on September 29, and Anthropic announced Claude Opus 5.5 on September 22. The choice suggests Meituan is making a sub-frontier case — performance competitive with an earlier generation of premium models on selected tasks, at a much lower token price. That is a reading of the evidence Meituan chose to publish, and the choice alone cannot establish how LongCat ranks against today’s best models.
For a software team, the split across benchmarks matters because terminal operations, repository fixes and translating a screenshot into a web interface are different jobs. A practical comparison should hold the tools and budget constant: give each model the same repository state, browser, test suite, tool permissions and time limit, repeat a few tasks, and score the rendered result and the code separately. A polished answer can hide a failed tool call or a change that breaks an unrelated feature.
Pricing and the MIT License: Cheap to Try, Free to Run
Meituan’s hosted rate card is aggressive. The limited-time rate is $0.30 per million uncached input tokens and $1.20 per million output tokens, with cached input at $0.006 per million. At those rates, a run with 200,000 uncached input tokens and 100,000 billable output tokens would cost $0.18 in model tokens. The same token counts would cost $0.09 at DeepSeek’s off-peak rates, $1.40 with GPT-6.1 Sol, $2.80 with Claude Opus 5.5, $7 with GPT-6 Astra and $5.60 with Claude Fable 5.2. LongCat retains a sizable price advantage even against the cheaper successors of the frontier models — though in practice, models can spend different amounts on reasoning, repeat tool calls or require a second attempt, so the honest metric is cost per accepted result, including retries and the time engineers spend reviewing patches.
The MIT license is the other half of the story. It allows commercial use, modification and redistribution subject to its notice requirements, with one reservation: it grants no rights over Meituan’s trademarks or patents. For teams that were waiting on license terms before touching a large open-weight model, this is the most permissive license short of public domain. It also means the model can run inside private infrastructure, which is exactly the deployment some regulated teams need — and it is the same direction the wider market is moving, from AMD’s new EPYC 6 Venice server CPUs, built for the AI agent era to ElevenLabs’ move beyond voice into enterprise agents.
The Catch: 3.5 Terabytes to Self-Host
Downloading the weights is the easy part; running them is not. With roughly 1.6 trillion total parameters and a full checkpoint required at serving time, self-hosting is a multi-terabyte, multi-GPU undertaking. At two bytes per parameter, 1.6 trillion parameters would imply about 3.2 TB for weights alone in a hypothetical uniform 16-bit representation — and that leaves out runtime buffers and the memory used to retain context across a request.
Meituan points to an SGLang-FluentLLM deployment guide for neural processing units, and its technical blog describes reducing the precision of expert weights and activations to four bits on Chinese accelerators. Those optimizations are evidence about Meituan’s tested setup. For teams with another accelerator fleet, the serving runtime has to be validated: can it load the checkpoint and preserve the model’s visual and tool-use behavior, how many long requests can the cluster handle concurrently, and how does performance change under load? A private deployment is worth investigating when the visual coding workload is large or sensitive enough to justify operating that stack. A team running occasional prototypes has a simpler path through a hosted evaluation — try it first at longcat.ai or through the API before buying a cluster.
The Washington Question
One important note for U.S. and Western enterprises considering this model: Chinese AI models are facing growing scrutiny in Washington. In October, the U.S. Department of Homeland Security disclosed an order directing federal civilian agencies to exclude DeepSeek from procurement and remove its technology from covered systems, including implementations supplied by third parties. VentureBeat found no publicly disclosed U.S. federal procurement or use restriction that specifically names Meituan or LongCat-2.5 as of November 9 — but the rules can change quickly.
For enterprises considering LongCat-2.5, especially those that serve or plan to serve U.S. government clients, where the model runs and whether a contract permits its use are separate questions. Self-hosting, or running the weights in an enterprise-controlled cloud, can give teams more control over where prompts and logs go, which model version they run and what tools the model can use. A restriction can still follow the technology into a third-party deployment, as the DeepSeek order illustrates. Suppliers should check their contract terms and applicable federal supply-chain orders before putting LongCat-2.5 into government-facing work — and keep a second model tested on the same tasks, so a change in eligibility does not force a rebuild.
How to Try It: Three Tests Before You Commit
The first decision is whether LongCat-2.5 improves a workflow the team already has. Start with three tests: a screenshot-driven bug fix, a design-to-code task, and a text-only repository task as a control. If it gets more of the visual work through review at a lower all-in cost, route that work to it; a cluster purchase can wait until the trial establishes how often that advantage shows up and what it is worth.
For the hosted trial, use a sandbox with public or synthetic artifacts first, and establish how the service handles retention, access and deletion before uploading company data. Cache rates matter for agents that repeatedly send a stable prompt prefix or large amounts of repository context — measure the actual hit rate in the tool developers use, because the size of the context window alone tells you little about the bill. And for text-heavy maintenance work, DeepSeek-V4.1-Flash is a useful first comparator, given its published coding scores and lower off-peak price; for changes where reviewer time or the consequences of failure dominate the bill, keep a stronger performer available as a fallback.
Conclusion
LongCat-2.5 is the most interesting open-weight release of the month, not because it beats the frontier models — on the newest benchmarks it does not — but because of what it makes cheap. A trillion-parameter vision-and-code agent under the MIT license, with a hosted API at $0.30 per million input tokens, turns screenshot-driven coding from a premium-model luxury into something a small team can test this week. The self-hosting bar is real, the benchmark claims are Meituan’s own, and the policy picture for Chinese models in U.S. procurement is unsettled. But the direction is clear: agent models are getting bigger, cheaper and more open at the same time, and LongCat-2.5 is the latest proof that the open-weight route can now compete where it matters — on the work, not just the weights.
Related reading: Anthropic’s Claude spreadsheets, meetings and whiteboards show the office side of the same agent race.
Sources: VentureBeat on the LongCat-2.5 release · Silicon Report on the weights release · Neural Daily on LongCat-2.5
