DeepSeek V4-Pro Is GA: What an MIT-Licensed Frontier Model Actually Changes

DeepSeek V4 is DeepSeek’s mixture-of-experts model family. The V4 preview, covering V4-Pro and V4-Flash, launched on 24 April 2026. DeepSeek-V4-Pro reached general availability as DeepSeek-V4-Pro-0813 on 13 August 2026.

The weights and repository are published under the MIT licence on Hugging Face at deepseek-ai/DeepSeek-V4-Pro-0813. That covers the weights and code, not the training data, so “open weight” is the more precise term than “open source”.

As of September 2026: $0.66 per million input tokens off-peak and $1.32 at peak on a cache miss, $0.022 and $0.044 on a cache hit, and $1.98 and $3.96 per million output tokens. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays.

Yes, and DeepSeek publishes vLLM and SGLang serving recipes. Be realistic about scale: it is a 1.7T-parameter model and the reference configuration in the model card is a multi-GPU GB300 node.

TrueFoundry’s AI Gateway adds roughly 3-4 ms and sustains 350+ RPS on 1 vCPU, which is immaterial next to multi-second generation times.

No, and the benchmark spread above is the argument against it. Routing by task type across open and closed models is usually cheaper and better than standardising on one.