Meituan Releases LongCat-2.5 Open Weights Under MIT Licence — With a 1.6T vs 1.8T Puzzle

Meituan has released the weights for LongCat-2.5, its 1.6-trillion-parameter coding model, under the MIT licence — but the model's own Hugging Face page cannot agree on how big the model is, with the card body saying 1.6 trillion parameters while the repository metadata displays 1.8 trillion. The weights were published on 9 November 2026, one week after the model was introduced, according to VentureBeat, and the release immediately became one of the most talked-about open-weights drops of the year. "Open weights" is a precise term with a precise meaning, and it is not the same as open source — a distinction that matters a lot here, because what Meituan has published is the model itself, not the recipe for making it.

What LongCat-2.5 is

LongCat-2.5 is, in the model card's own words, "a native multimodal MoE model for long-horizon agentic tasks". MoE stands for mixture of experts, an architecture in which only a subset of the model's parameters is active for any given token — here about 48 billion parameters activated per token out of 1.6 trillion in total. Native multimodal means the model was built to handle visual tokens jointly aligned with text from the start, rather than having vision bolted on afterwards. The model's stated purpose is long-horizon agentic coding: working through complex codebases and full-project builds over extended sessions, rather than answering single questions or completing short snippets. It supports a 1-million-token context window, verified in the API documentation as 1,048,576 tokens, which is what lets it hold a large project in view at once.

The model was announced on 1 November 2026 on Meituan's tech blog, with the full API release following on 2 November, after a preview on 25 September. The weights release on 9 November is the third step in that sequence, and it is the one that matters most to the open-model community, because it lets anyone download and run the model rather than renting API access.

The 1.6T versus 1.8T puzzle

Here is the discrepancy, stated plainly. The Hugging Face model card body says LongCat-2.5 has 1.6 trillion total parameters. The repository metadata on the same Hugging Face page displays 1.8 trillion. And the weight index inside the repository records 1.78 trillion parameters — 1,779.8 billion, to be exact. Three numbers, one model, no official explanation at the time of writing. This article does not pick one figure silently and does not present any single number as settled. One possible explanation circulating online is that the difference comes from embedding tables being counted in one total and not in another, but that is speculation, not a Meituan statement, and it is not presented here as fact. Until Meituan clarifies, the honest description is that the card says 1.6 trillion while the repository records about 1.78 trillion, and readers checking the page themselves will see both.

A similar caution applies to the model's size on disk. The checkpoint has been reported at about 3.56 TB in BF16 format, across 208 files — but that figure comes from a single source reporting the repository size, not from Meituan itself. VentureBeat instead calculated about 3.2 TB as illustrative arithmetic, multiplying 1.6 trillion parameters by two bytes each, and explicitly said that is not the actual footprint. So the accurate way to describe it is a multi-terabyte checkpoint, over 3 TB, reported at about 3.56 TB in BF16.

A programmer working late at a desk with multiple monitors showing code repositories.
A programmer working late at a desk with multiple monitors showing code repositories.

How good is it? Read the asterisks

Meituan's headline benchmark is Terminal-Bench version 2.1, a test of agentic coding ability. The model card gives LongCat-2.5 a score of 88.4, against 89.1 for Claude Opus 5 and 88.8 for GPT-5.6-Sol — close enough to put it in the same conversation as the leading proprietary models. But the fine print matters more than the headline. The card itself explains that scores marked with an asterisk are cited from public reports or leaderboards, while unmarked scores are "measured in-house and are indicative only". LongCat-2.5's 88.4 is unmarked: it is Meituan's own in-house measurement. The Opus 5 score of 89.1 and the GPT-5.6-Sol score of 88.8 both carry asterisks, meaning they were taken from public sources. All three numbers must therefore be labelled vendor-reported or in-house and indicative only — they are not independent, head-to-head results from a neutral tester, and they should not be quoted as if they were.

That caveat does not make the score meaningless. A model that its own maker measures at 88.4 on a demanding agentic benchmark is clearly a serious coding model, and the gap to the best proprietary systems is small on this one test. But one benchmark, measured in-house, is not a verdict. Developers should treat the 88.4 as a reason to test it, not as a guarantee.

Open weights is not open source

The MIT licence on the weights is genuinely permissive: the card states plainly that "the model weights are released under the MIT License", which allows commercial use, modification and redistribution. One limit is worth noting: the licence does not grant rights to use Meituan's trademarks or patents. But the more important point is what the release does not include. Meituan has published the weights plus serving and deployment code — including a guide for deploying with SGLang-FluentLLM on NPUs and W4A4 quantisation support — but it has not published the training data or the full training pipeline. That is the standard definition of "open weights" as opposed to "open source": you can download, run, fine-tune and deploy the model, but you cannot reproduce how it was built or audit what it was trained on.

Can you actually run it?

Realistically, almost nobody will self-host LongCat-2.5 on a workstation. A multi-terabyte checkpoint needs data-centre hardware, and even with W4A4 quantisation the model is far beyond what a single machine can run. The realistic route is Meituan's hosted API, priced at $0.30 per million input tokens and $1.20 per million output tokens, with cached input at $0.006 per million tokens. Those are described as a limited-time price, and Meituan publishes no list price and no end date for the promotion, so the hosted price could change without warning. The Hugging Face card itself gives no prices at all — pricing comes from the API documentation, not the model page.

One more figure deserves a caveat. The maximum output length is disputed between Meituan's own pages: the model-parameters page lists 262,144 tokens, while the quick-start page lists 128K. This article states both rather than choosing one. Developers building on the API should check the live documentation for the limit that applies to their setup.

Close-up of a terminal screen showing code with a coffee cup beside the keyboard.
Close-up of a terminal screen showing code with a coffee cup beside the keyboard.

Why this release matters

LongCat-2.5 matters for two reasons. First, it is a genuinely capable agentic coding model released under a permissive licence, and its in-house benchmark score puts it near the frontier on at least one demanding test. That gives developers and small teams a frontier-class coding model they can run and fine-tune without depending on a proprietary API. Second, it comes from Meituan, a Chinese food-delivery and local-services giant rather than a company known for AI research, which underlines how far the open-weights movement has spread beyond the usual names. Readers tracking the open-model ecosystem may also want to revisit our coverage of Mistral Large 4, another major open-weights release, and the Impeccable GitHub skill and Hallmark GitHub skill, two high-star projects from the same GitHub ecosystem where LongCat-2.5 is now being discussed.

A server room corridor with rows of server racks and blinking lights.
A server room corridor with rows of server racks and blinking lights.

Key Takeaways

  • Meituan released LongCat-2.5 weights on 9 November 2026 under the MIT licence, a week after introducing the model — the weights, not the training data or pipeline, so this is open weights, not open source.
  • The parameter count is genuinely disputed: the card body says 1.6T, the repo metadata shows 1.8T, and the weight index records 1.78T — do not quote one figure as settled.
  • Terminal-Bench v2.1 gives LongCat-2.5 an 88.4 in-house score versus 89.1 for Claude Opus 5 and 88.8 for GPT-5.6-Sol — all vendor-reported or cited, indicative only, not independent results.
  • The hosted API costs $0.30/$1.20 per million tokens as a limited-time price, with no published list price or end date; the max output limit is disputed (262,144 vs 128K tokens).
  • Self-hosting is unrealistic for most users: the checkpoint is multi-terabyte, reported at about 3.56 TB in BF16 — a reported figure, not a Meituan one.

Conclusion

LongCat-2.5 is a significant open-weights release: a native multimodal, long-horizon coding model with a permissive MIT licence and an in-house benchmark score near the best proprietary systems. But the release is also a case study in reading the fine print. The parameter count differs across Meituan's own pages, the benchmark scores are in-house or cited rather than independent, and the hosted price is a limited-time offer with no end date. Download the weights, test the model on your own code, and quote the numbers with their caveats — that is the honest way to use a release this interesting.

Sources

Comments

Post a Comment