Open LongCat-2.5 Weights Released

Meituan has released the open weights for its LongCat-2.5 vision-language model under the MIT License, one week after the model was introduced. According to VentureBeat, the company placed the checkpoint on Hugging Face and said it gave developers another way to run a model that can interpret images and video, navigate codebases, and generate or update software across multiple steps. The release matters because it adds a new access route for developers, even as the source noted that questions about how the model performs in enterprise settings remain.

What Meituan Released and When

Meituan said it released the weights for LongCat-2.5 on Nov. 9. The timing was one week after the model’s introduction. The license was described as permissive. The source reported the checkpoint appeared on Hugging Face. Meituan’s technical material described visual information as tokens and said the model processed visual input alongside text. Its API guide showed how developers could submit screenshots, charts, images and videos, with images embedded in requests or supplied by URL. The video interface took a URL and a meaningful constraint for teams using internal screen recordings, the source said.

Why Visual Coding Is the Key Test

The strongest case for testing LongCat-2.5 is a job that requires both visual judgment and coding. According to VentureBeat, the test is a developer giving an agent a screenshot of a broken checkout flow, asking it to locate relevant components in a repository, propose a patch and check the result in a browser. Meituan’s hosted API pricing made that experiment relatively inexpensive, the source said. The more useful question for buyers was whether it could handle enough visual and coding work at discounted API rates to reserve premium models for harder jobs. Operating the full model inside a private network presented a different calculation.

For frontend and product engineering, Meituan pointed to an initial task such as a visual regression with a known fix. A team could give the agent a failing screenshot, access to a test repository and a way to render the patched page, then check whether it found the cause, made a targeted change and preserved keyboard navigation and responsive behavior. Meituan also said a similar workflow could turn design references into prototypes, but the test should check whether output followed the team’s component library and interaction rules as well as whether it looked right.

How LongCat-2.5 Compares

Meituan’s clearest comparison was with its predecessor. LongCat-2.5 scored 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, according to Meituan’s evaluation. Rival comparisons in the launch material were more mixed. LongCat-2.5 was within one point of GPT-5.6 Sol and Claude Opus 5 on Terminal Bench v2.1. On version 4.0, it scored 38.5 versus Claude Opus 5’s 53.9. DeepSeek-V4.1-Flash led the four models below on DeepSWE, while LongCat led that group on 3DCodeBench and Claude Opus 5 led on Vision2Web.

Meituan chose GPT-5.6 Sol and Claude Opus 5 as 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 suggested Meituan was 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. The source said that is a reading of the evidence Meituan chose to publish, and that the choice alone cannot establish how LongCat ranks against today’s best models. Newer Terminal-Bench results changed 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 posted higher leaderboard scores than LongCat’s reported 38.5, though differences in setup prevented a clean head-to-head conclusion.

What the License, Pricing and Deployment Mean

The MIT license on Meituan’s weights allows commercial use, modification and redistribution, subject to notice requirements, according to the source. The hosted service is a separate purchase decision. Meituan’s rate card listed a limited-time rate of $0.006 per million cached input tokens, plus $0.30 for uncached input and $1.20 for output. The source gave an example: 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, the source said.

Self-hosting depends on the serving stack. According to Meituan’s model card, LongCat-2.5 has about 48 billion active parameters per token out of 1.6 trillion total. The model routes each token through a subset of its experts, but serving still requires making the full checkpoint available, the source said. 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, though the source called that arithmetic illustrative. The Hugging Face repository snapshot displayed 1.8T in metadata, while Meituan’s technical description said 1.6T. Meituan pointed to an SGLang-FluentLLM deployment guide for neural processing units, and its technical blog described reducing precision of expert weights and activations to four bits on Chinese accelerators.

How This Fits With Grafy

For creators and developers, the practical connection to Grafy is the coding-agent workflow. The source’s strongest use case was still visual coding: giving an agent a screenshot, repository context and browser check. In Grafy, the relevant app is Grafy Code, an AI coding studio that installs packages and runs code. A reader could also use Grafy Chat to compare multiple models. The story does not mean Grafy has a LongCat-specific feature; it means the broader challenge of multimodal coding evaluation is relevant to Grafy’s audience.

Frequently asked questions

Does Meituan’s MIT License Permit Commercial Use of LongCat-2.5?

According to the disclosed information, the MIT license on Meituan’s weights allows commercial use, modification and redistribution, subject to its notice requirements. The hosted service is described as a separate purchase decision. The source did not spell out every compliance consequence, so organizations should review the license and service terms before relying on the model for commercial products.

Does LongCat-2.5 Teach a Smaller Model?

The source material does not say that LongCat-2.5 teaches a smaller model. Meituan’s reports focus on LongCat-2.5’s capabilities, benchmarks and release of its weights. The disclosed information does not present a smaller-model teaching claim, so this cannot be treated as an established finding.

Was the LongCat-2.5 Weights Release the Same Version Meituan Introduced?

Meituan introduced LongCat-2.5 and released the weights one week later, on Nov. 9. The source reports that timing, but it does not clearly state whether the released checkpoint was identical to the introduced version. Meituan said the checkpoint was published on Hugging Face under the MIT License.

---

Source: VentureBeat — AI — Meituan releases the open weights of its LongCat-2.5 LLM with vision under a permissive MIT License