DeepSeek published the weights for V4-Flash-Vision-Exp on Hugging Face on 31 August. The repository timestamps come straight from the platform's API: created 06:16:18 UTC, shards uploaded from 06:57 UTC, model card at 07:57 UTC. That is 14:16 CST — Monday afternoon in China.
Why the date is the story
DeepSeek announced this model on 21 August, and its own documentation described it as available through the API. Weights were not mentioned. Model trackers accordingly date it to 21 August, so today's release is being logged as a duplicate of something ten days old rather than as the thing that actually happened — a closed model becoming an open-weights one. For anyone who cannot send data to an API, those are not the same event, and only one of them changes what they can deploy.
What is in the release
MIT licence. 48 shards, 167,811,372,792 bytes — about 167.8 GB. 43 layers, 256 routed experts plus one shared, six active per token, a 1,048,576-token context, FP8 e4m3 weights with experts at FP4, and a native 32-layer vision tower. An MIT licence on a mixture-of-experts multimodal model at this scale is unusually permissive — no acceptable-use appendix, no revenue threshold, no review clause of the kind that has been appearing in Chinese open-weight licences recently.
What the common framing gets wrong
The multimodal headline number deserves care. The flagship gain quoted for this model is measured against a baseline that DeepSeek's own footnote says does not receive the images — so part of the improvement is scored against a text-only comparison, which is a floor rather than a competitor. The card is honest about this; the summaries repeating the number are not. Two other things get dropped: at least one coding-security benchmark regressed against the previous version, and the model loses most head-to-head rows against the leading Western frontier model. "Comparable text performance" is doing a lot of averaging.
One claim that did not hold
Coverage describes a simultaneous release on Hugging Face and ModelScope. ModelScope returns record not found for the model. As of publication the weights are on Hugging Face only.