DeepSeek V4 Pro exits preview with MIT-licensed 1.7T weights

DeepSeek's flagship went from preview to production this week with the kind of rollout that would be unusual for any other AI lab. The model card on Hugging Face went live, the benchmark table reportedly leaked out of an official WeChat group and then a Reddit post that later disappeared, and, as Simon Willison noted on his blog, 'DeepSeek don't have any obvious announcement page for their new model.'

V4 Pro is a mixture-of-experts system with roughly 1.6 trillion total parameters and 49 billion active per token; the weights posted to Hugging Face weigh in at 1.7T and 893 GB. It ships under an MIT license, keeps DeepSeek's now-standard million-token context window, and adds a speculative decoding module the company calls DSpark. The card claims the release 'outperforms DeepSeek-V4-Pro (Preview) on listed benchmarks and is broadly competitive with the strongest proprietary models available,' pointing to numbers like 87.9 on Terminal Bench 2.1, 60.0 on HLE with tools, and 83.3 on Cybergym.

That MIT tag is the part that will keep landing. A frontier-scale model is now downloadable, redistributable, and commercially usable by any cloud provider that can rack the hardware to serve it. According to Unite.AI, the preview period this replaces ran about four months, and DeepSeek has also warned of 'a significant increase' in its API pricing, which makes the open-weights route a lot more interesting for anyone who was leaning on the hosted API as their cheap-inference plan.

A few things to sit with before treating the numbers as gospel. The benchmarks came out via WeChat, then a deleted Reddit post, before landing as an ASCII table on Hacker News, per Willison; there is no press page and no independent replication yet. The card also does not name which proprietary models it means by 'broadly competitive,' so that phrase is doing quiet work. And at 893 GB across BF16 and FP8 tensors, this is open in the licensing sense, not the laptop sense: serious expert-parallel infrastructure is the entry ticket.

Three of the AI experts in our Who's Who directory already shared the source link, which is the tell that this one is going to get poked at hard over the coming days. If the numbers hold on independent runs, the interesting move is at the inference-provider layer: whoever gets a well-tuned V4 Pro endpoint up first, at a price that undercuts DeepSeek's own coming hike, has an argument to make.