Meituan releases LongCat-2.5 weights on Hugging Face under MIT licence

Meituan has published the weights of LongCat-2.5, its 1.6-trillion-parameter multimodal model for coding agents, on Hugging Face under the MIT licence. When Unicorn API covered the launch, the weights were marked “coming soon” on both GitHub and Hugging Face and the licence was unconfirmed. Both questions now have answers: the checkpoint is downloadable, and the licence is one of the most permissive in common use.

What changed since the launch report

Meituan’s LongCat team released LongCat-2.5 on November 1, positioning it as a native multimodal model for long-running agent tasks. At that point developers could only reach it through Meituan’s hosted API.

According to Grafy, which cites VentureBeat, Meituan said it released the weights on November 9, one week after the model’s introduction, and placed the checkpoint on Hugging Face. Meituan framed it as another way for developers to run a model that can interpret images and video, navigate codebases, and generate or update software across multiple steps. Developer posts on November 12 are now pointing builders to the repository.

The repository is published under the meituan-longcat organisation on Hugging Face and includes the tokenizer configuration and chat template. The model itself is unchanged from launch: the news is access, not new capabilities.

What is in the LongCat-2.5 release

LongCat-2.5 is a mixture-of-experts model: a design in which each token is routed through only a subset of the network’s “experts”, so the active compute per token is much smaller than the total size. Meituan’s figures put it at 1.6 trillion total parameters with about 48 billion active per token. Apiyi notes that some launch-day community posts listed the active figure as 440B, while Meituan’s official blog says about 48B.

Spec	LongCat-2.5
Total parameters	1.6T (Meituan); repository metadata showed 1.8T
Active parameters per token	About 48B
Context window / maximum output	1M tokens / 128K tokens
Inputs	Text, image URLs, video URLs
Hosted API formats	OpenAI and Anthropic formats
Licence	MIT

One discrepancy is worth noting: Grafy reports that the Hugging Face repository snapshot displayed 1.8T in its metadata, while Meituan’s technical description says 1.6T. The sources do not explain the difference.

The chat template shows how the model expects input. Image and video items are wrapped in dedicated start and end tokens, and both are passed as URLs. The template also carries a thinking switch that defaults to off, a tool-choice setting that defaults to auto, and support for tool calls and tool results in conversation history. A built-in code interpreter tool is declared as a network-isolated, stateful Jupyter sandbox that times out after 60 seconds by default and accepts Python, Java, Go, JavaScript, TypeScript, C and C++.

“Native multimodality” has a specific meaning in Meituan’s description. Images are converted into discrete tokens at native resolution using a semantic-aligned encoder and residual vector quantization, and the backbone processes visual and text tokens in one format rather than splicing in features from a separate vision encoder. Because the tokenizer is fixed, visual token IDs can be precomputed and cached offline.

What the MIT licence allows

According to Grafy, the MIT licence on Meituan’s weights allows commercial use, modification and redistribution, subject to notice requirements. That puts LongCat-2.5 among the open-weight models with no usage-based restrictions described in the sources, which matters for teams that want to fine-tune, embed or redistribute a model in a commercial product.

The hosted service remains a separate decision. Grafy describes Meituan’s API as its own purchase option with a limited-time discounted rate card, distinct from running the weights yourself. Before the open release, LongCat 2.5 had also appeared as a free preview: Sulat’s model catalogue lists a “LongCat 2.5 Preview Free” through OpenCode with a release date of September 25, 2026, a 1,000,000-token context window and a 131,072-token output cap, now marked deprecated.

Running a 1.6T model yourself

Open weights do not make LongCat-2.5 small. The model routes each token through a subset of its experts, but serving still requires the full checkpoint to be available, Grafy notes. At two bytes per parameter, 1.6 trillion parameters would imply about 3.2 TB for the weights alone in a uniform 16-bit format; Grafy calls that arithmetic illustrative.

Meituan’s own engineering points to lower-precision serving. It refers developers to an SGLang-FluentLLM deployment guide for neural processing units, and its technical blog describes reducing expert weights and activations to four bits on Chinese accelerators. Apiyi’s summary of the official blog calls this W4A4, meaning both weights and activations are quantized to 4 bits. In Meituan’s online load test against a W8A8 INT8 baseline, it reports:

• prefill throughput up 43.3%
• decode throughput up 42.77%
• time to first token down 15.91%
• time per output token down 19.72%

Those are Meituan’s measurements on its own stack. Training ran on domestically produced hardware with less than 80GB of memory per GPU, using techniques such as double-sharded FSDP, hybrid vocabulary parallelism and dynamic expert replication, according to the official blog as summarised by Apiyi. The sources do not report independent throughput or memory figures on common GPU clusters.

Grafy also flags a practical constraint: the video interface takes a URL, which matters for teams that want to feed in internal screen recordings that are not publicly hosted.

How Meituan’s benchmark claims read now

Every score published so far comes from Meituan. Against its predecessor, Meituan reports large gains on several agentic coding tests. In Meituan’s evaluation as reported by Grafy, 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. Apiyi’s summary of Meituan’s results adds Terminal Bench v4.0 rising from 3.0 to 38.5 and Program Bench Almost Resolved rising from 2.5 to 25.5, with AutomationBench v1.0.6 also climbing sharply from 15.8 for LongCat-2.0.

Against rivals, the picture Meituan published is more mixed:

• On Terminal Bench v2.1, LongCat-2.5 was within one point of GPT-5.6 Sol and Claude Opus 5.
• On Terminal Bench v4.0, it scored 38.5 versus Claude Opus 5’s 53.9.
• In a group of four models in Meituan’s table, DeepSeek-V4.1-Flash led on DeepSWE, LongCat-2.5 led on 3DCodeBench, and Claude Opus 5 led on Vision2Web.

Grafy points out that Meituan’s proprietary reference models already had successors at launch: OpenAI released GPT-6.1 Sol on September 29, and Anthropic announced Claude Opus 5.5 on September 22. Its reading is that Meituan is making a sub-frontier case: competitive with an earlier generation of premium models on selected tasks, at a much lower token price. Newer Terminal-Bench leaderboard results show both GPT-6.1 Sol and Opus 5.5 above LongCat’s reported 38.5, though Grafy says setup differences prevent a clean head-to-head conclusion.

The open release should help close this gap. With weights public, third parties can now run their own evaluations rather than relying on the maker’s tables.

What it means for builders

The clearest use case in the sources is work that mixes visual judgment with coding. Grafy, citing VentureBeat, describes the test as giving an agent a screenshot of a broken checkout flow, asking it to locate the relevant components in a repository, propose a patch and check the result in a browser. Meituan suggests starting with a visual regression that has a known fix, then checking whether the model found the cause, made a targeted change and preserved keyboard navigation and responsive behaviour. For design-to-prototype work, the check should include whether output follows the team’s component library and interaction rules, not only whether it looks right.

Developers are proposing similar tests. One suggests giving it an existing settings page and a revised UI design, then asking it to update the layout without breaking form validation or save functionality.

Meituan’s training choices target long agent runs. The team says it turned real agent failure trajectories into reproducible, verifiable tasks, and added penalties for repeated reasoning and repeated tool calls to curb overthinking. If that holds up in practice, it would reduce the token waste common when agents loop on the same tool.

Practical points for teams weighing the weights:

• Licence: MIT permits commercial use, modification and redistribution with notice.
• Hardware: the full checkpoint must be served; Meituan’s tooling is oriented to four-bit serving on neural processing units.
• Thinking mode: the chat template defaults reasoning to off, so it must be enabled explicitly.
• Evidence: all capability claims remain Meituan’s own.

What to watch next

The main open question is independent evidence. Third-party benchmark runs on the released weights, especially on Terminal Bench v4.0 and the visual coding tests where Meituan’s own numbers are mixed, would show where LongCat-2.5 actually sits against current models. Also worth tracking: whether Meituan clarifies the 1.6T versus 1.8T parameter discrepancy in the repository metadata, whether deployment guides appear for GPU stacks beyond its neural-processing-unit setup, and whether quantized community builds bring the memory footprint within reach of smaller teams.