DeepSeek R3 dropped at midnight Beijing time. The open-weights story of the quarter starts here.

It dropped at 00:11 Beijing time. No press release. No tweet. Just a commit message that read "v3.0 weights, see LICENSE" and a model card with one paragraph of English and one of Mandarin.

DeepSeek R3 is the open-weights story of the quarter, and it isn't because of the architecture. The architecture is not new — it's a refinement of the v2-Lite mixture-of-experts design that made R2 the most-downloaded weights of 2025. The story is the LICENSE file.

What the license actually says

You'll notice the new license is not Apache. It is also not the Llama "permissive plus carve-outs" template. It is something the maintainers are calling DeepSeek-OW-1.0, and it has three clauses worth reading word by word: a downstream use restriction (clinical decision support requires a separate paid tier), an attribution requirement (any deployed system must surface the model name and version on user-facing surfaces), and a "compute-floor" clause that exempts academic and individual users from the attribution requirement under a defined parameter-count threshold.

It is, in the precise meaning of the word, "open weights, restricted use." It is also the most legally defensible such license a Chinese lab has published. Counsel at three U.S. companies I spoke to over the weekend has read the terms and is not blocking adoption.

What developers will care about

• The v3 weights load on existing R2 inference stacks with a one-line config change (router_topk: 8 → 12).
• Quantized 4-bit GGUFs were posted by community members within four hours of release.
• The fine-tuning kit, including LoRA adapters, lands in the next 48 hours according to the maintainers.

The benchmarks are the part of the conversation we'll have in two weeks, when the dust settles and someone runs Mètis-v0.3 on it. For this morning, the news is the license. Read it before you build on it.