LongCat 2.5 Weights Out: MIT License, 3.3 TB Download

The most repeated line in LongCat 2.5's launch week was "weights coming soon" — the official blog said it on Nov 1, every report from Nov 1–3 repeated it, and even a three-way comparison updated on Nov 4 still called the model API-only. But that line expired on Nov 4, when two weights repositories quietly went live on Hugging Face under the MIT licence, the full release at 3.3 TB. ModelBrief verified both repositories' creation times, licences and sizes via the Hugging Face API; this article uses the verified figures.

Timeline: from "coming soon" to a quiet listing

On Nov 1, Meituan published the LongCat 2.5 technical blog with weights listed as coming soon; on Nov 2 the API platform went GA, with video input and Codex access added, but still no weights. English analyses and Chinese reports from Nov 1–3 — including eesel's, the most thorough third-party read — all proceeded on the assumption that weights were unpublished. The turn came on Nov 4: the full repository and the W4A4 quantized repository were created on the same day, with no separate announcement. The news travelled far slower than the repositories: ToolPilotPro's comparison, updated Nov 4, still called LongCat API-only, and its verdict of "the only model you can't own" kept being quoted everywhere. When ModelBrief checked on Nov 13, nine days after release, both repositories still had only three-digit download counts — many people still don't know the weights are out.

The two repositories, in facts

Download counts are as checked on Nov 13. Both repositories confirm the MIT licence twice over — a LICENSE file and a README badge; a ModelScope mirror also exists. The contents are thorough: config.json confirms 38 layers, 768 routed experts, top-k 12, MLA attention and a 163,840-token vocabulary; a chat_template.jinja is included; the README carries the full evaluation table, identical to the official blog, plus a quantization section.

What MIT means

MIT is one of the most permissive mainstream open-source licences: commercial use, modification and redistribution are all allowed, with almost no attached conditions. For LongCat 2.5, that makes the weights release a real handover of the model, not a gesture. For comparison: DeepSeek also ships MIT weights, putting the two on a par for openness; Kimi K3.1 Pro, by contrast, uses a custom licence with a $20M threshold for MaaS (model-as-a-service) — read the licence terms before any commercial deployment. Before Nov 4, "LongCat is less open than Kimi" was true; after Nov 4, it is the other way round.

Deployment reality: downloadable is not the same as deployable

The weights release changes the openness verdict in the comparison, but the self-hosting reality stays demanding. First, the size: 3.3 TB full, 1.24 TB quantized — that alone means datacentre-grade deployment, not a personal workstation. Meituan has published no official VRAM or GPU configuration guidance, and ModelBrief does not estimate it — any claim of a specific GPU count should be treated with suspicion. Second, the deployment path: the README points only to NPU platforms and Meituan's own SGLang-FluentLLM fork; whether vLLM or mainline SGLang supports 2.5 has no official statement (unverified). In other words, to run it, you first have to accept the NPU route Meituan prescribes.

Against 2.0, this was actually fast: 2.0 was likewise MIT, GA on June 30, with its Hugging Face repository created July 5; 2.5 went from Nov 2 GA to weights on Nov 4 — just two days. Meituan didn't delay open-sourcing; what lagged was the spread of the news. For model selection, the conclusion cuts both ways: the weights are out, so LongCat is no longer a "rent-only" model; but the 3.3 TB size plus the NPU orientation means most users will still go through the API — open-sourcing changes bargaining power and long-term options, not tomorrow's deployment.