Meituan releases the open weights of its LongCat-2.5 LLM with vision under a permissive MIT License

Meituan released the weights for LongCat-2.5 on Nov. 9, 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 another way to run a model that can inspect screenshots and video, work through a codebase, and write or revise software over multiple steps.

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.

Meituan’s hosted API pricing makes that experiment relatively inexpensive. The more useful question for buyers is whether it can handle enough visual and coding work at its discounted API rates to reserve premium models for harder jobs. Operating the full model inside a private network presents a different calculation.

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. The order targets DeepSeek and its affiliates.

VentureBeat found no publicly disclosed U.S. federal procurement or use restriction that specifically names Meituan or LongCat-2.5 as of Nov. 9. But the rules can change quickly.

As VentureBeat reported in February, Defense Secretary Pete Hegseth announced on Feb. 27 that he was directing the Pentagon to designate Anthropic a “Supply-Chain Risk to National Security.” He formalized the designation on March 3. The two cases show how quickly a dispute or a supply-chain decision can change what agencies and their suppliers can use

Where visual understanding earns its keep

In its technical write-up, Meituan describes a model that represents visual information as tokens and processes it alongside text. Its API guide shows how developers can submit screenshots, charts, images, and videos. Images can be embedded in a request or supplied by URL. The documented video interface takes a URL, a meaningful constraint for teams working with internal screen recordings.

For frontend and product engineering, a good initial task 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. Check whether it finds the cause, makes a targeted change, and preserves keyboard navigation and responsive behavior. A similar workflow can turn design references into prototypes, but the test should check whether the output follows the team’s component library and interaction rules as well as whether it looks right.

Meituan highlighted developer demos that included a car-launch site from Tasihi, a fantasy game interface from AI Maverick Frank, and a shop management site from Aigc Lancet. The company’s own examples include an interactive fluid simulation and a 3D exhibition. These are good leads for teams building demos or internal tools. A storefront prototype and a production storefront have different acceptance criteria. For the latter, the generated code still has to pass the same functional, accessibility, and maintenance checks as a human-authored change.

LongCat gains on its predecessor while newer rivals set a higher bar

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 results make a case for retesting LongCat on repository work, particularly if teams previously ruled out 2.0.

The rival comparisons in Meituan’s launch material are more mixed. LongCat-2.5 comes within one point of GPT-5.6 Sol and Claude Opus 5 on Terminal Bench v2.1. On v4.0, it scores 38.5 versus Claude Opus 5’s 53.9. DeepSeek-V4.1-Flash leads the four models below on DeepSWE. LongCat leads this group on 3DCodeBench, while Claude Opus 5 leads on Vision2Web. For a software team, the split matters because terminal operations, repository fixes, and translating a screenshot into a web interface are different jobs.

Meituan chose GPT-5.6 Sol and Claude Opus 5 as two of its proprietary reference points. Their direct successors were already available when Meituan introduced LongCat-2.5. OpenAI released GPT-6.1 Sol on Sept. 29, and Anthropic announced Claude Opus 5.5 on Sept. 22. The choice of older baselines suggests Meituan is making a sub-frontier case for LongCat-2.5: 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. The choice alone cannot establish how LongCat ranks against today’s best models.

The table separates scores in Meituan’s benchmark report from scores on the Terminal-Bench 4.0 leaderboard. Meituan combines its own runs with published results. The leaderboard scores for GPT-6.1 Sol and Opus 5.5 use maximum effort and their respective coding tools, Codex and Claude Code. Treat them as reference points for the same benchmark version under different test setups. A dash means this comparison has no verified result to show. API prices are U.S. dollars per million uncached input and output tokens, respectively.

Meituan labels the benchmark entry DeepSeek-V4.1-Flash. DeepSeek’s API docs say its hosted Flash endpoint serves the 1017 checkpoint, but Meituan’s table leaves the checkpoint suffix unspecified.

The newer Terminal-Bench results change how to read the near tie between LongCat-2.5 and GPT-5.6 Sol in Meituan’s v4.0 table. GPT-6.1 Sol and Opus 5.5 both post higher leaderboard scores than LongCat’s reported 38.5, though the differences in setup prevent a clean head-to-head conclusion.

That makes the newer Sol and Opus the more useful proprietary controls for an enterprise coding trial. For teams that already pay for frontier performance on difficult work, GPT-6 Astra and Claude Fable 5.2, which Anthropic released Nov. 9, are escalation options to test on the cases LongCat struggles to complete.

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 to see how often each model succeeds.

For screenshot-to-code, score the rendered result and the code separately. For bug fixes, test whether the patch addresses the cause and whether the reviewer has to redo it. A polished answer can hide a failed tool call or a change that breaks an unrelated feature.

Pricing and licensing make a hosted trial easier

Meituan’s rate card lists a limited-time rate of $0.006 per million cached input tokens alongside $0.30 for uncached input and $1.20 for output. 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 Opus 5.5, $7 with GPT-6 Astra, and $5.60 with Fable 5.2 at the rates above.

The example holds token volume fixed and stays below the threshold where OpenAI charges higher rates for GPT-6.1 Sol. In practice, models can spend different amounts on reasoning, repeat tool calls, or require a second attempt.

The newer models also change the cost comparison in their own right. At base rates, GPT-6.1 Sol charges half as much for input and output as GPT-5.6 Sol’s listed promotional rate, and Opus 5.5 costs less per token than Opus 5. LongCat retains a sizable price advantage even against the cheaper successors.

DeepSeek’s off-peak price is a tighter benchmark for what an enterprise should expect to pay for a model with downloadable weights and competitive coding scores.

For purchasing decisions, track cost per accepted result. Include retries, browser or sandbox time, any fallback call to another model, and the time engineers spend reviewing patches.

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. The size of the context window alone tells you little about the bill.

The MIT license on Meituan’s weights allows commercial use, modification, and redistribution subject to its notice requirements.

The hosted service is a separate purchase decision. For a trial with proprietary code or screenshots, establish how the service handles retention, access, and deletion before uploading company data. A sandbox with public or synthetic artifacts is enough to learn whether the model merits a deeper review.

Self-hosting depends on the serving stack

According to Meituan’s model card, LongCat-2.5 has about 48 billion parameters active per token out of 1.6 trillion total. The model routes each token through a subset of its experts, but serving it still requires making the full checkpoint available.

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.

That arithmetic is illustrative. It leaves out runtime buffers and the memory used to retain context across a request, and it does not describe the actual checkpoint footprint.

The Hugging Face repository snapshot displays 1.8T in its metadata, while Meituan’s technical description says 1.6T. Use the files and the exact serving format to size a deployment.

Meituan points to an SGLang-FluentLLM deployment guide for neural processing units. 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, validate that the serving runtime can load the checkpoint and preserve the model’s visual and tool-use behavior. Then measure how many long requests the cluster can handle concurrently, how long those requests take, and how performance changes 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.

For text-heavy maintenance work, DeepSeek 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 on the relevant internal test available as a fallback.

The first decision is whether LongCat-2.5 improves a workflow the team already has. Start with 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 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 U.S. cloud can give teams more control over where prompts and logs go, which model version they run, and what tools it can use. A restriction can still follow the technology into a third-party deployment, as the DeepSeek order illustrates.

Even with no publicly disclosed order naming LongCat-2.5, suppliers should check their contract terms and applicable federal supply-chain orders before putting it into government-facing work. Keep a second model tested on the same tasks so a change in eligibility doesn’t force a rebuild.