DeepSeek publishes V4-Pro under an MIT licence at 1.7 trillion parameters
The model card reports 87.9 on Terminal Bench 2.1 and 83.3 on Cybergym, and ships with three selectable reasoning effort levels.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- DeepSeek-V4-Pro-0813 is published on Hugging Face under an MIT licence, among the most permissive terms any lab applies to weights at this scale.
- The card states 1.7 trillion parameters, speculative decoding, and low, high and max reasoning effort settings.
- Self-reported benchmarks include 87.9 on Terminal Bench 2.1, 83.3 on Cybergym and 62.7 on DeepSWE.
DeepSeek published DeepSeek-V4-Pro-0813 to Hugging Face on 13 August under an MIT licence. At a stated 1.7 trillion parameters it is among the largest models anyone has released weights for, and MIT is close to the most permissive licence available — no revenue share, no use restrictions, no acceptable-use rider.
The model card describes speculative decoding and three selectable reasoning effort levels, low, high and max, which shifts the cost-quality trade-off from a model choice to a per-request parameter. The card positions it as broadly competitive with the strongest proprietary models available, and as an improvement on the earlier preview release.
The benchmark table on the card reports 87.9 on Terminal Bench 2.1, 83.3 on Cybergym, 62.7 on DeepSWE, 74.1 on Toolathlon-Verified, 61.5 on NL2Repo and 60.0 on HLE with tools. These are self-reported, as benchmark tables on model cards always are — but unlike a hosted model, the weights are there for anyone who wants to check.
The strategic shape of this is worth noting separately from the numbers. Releasing frontier-adjacent weights under MIT sets a licensing floor that competitors either match or explain, and it does so at a parameter count that makes independent reproduction expensive enough to be slow.
An MIT licence at 1.7 trillion parameters puts the most permissive terms in the market on one of the largest open releases, which pressures every lab publishing weights under custom licences with revenue-share clauses. The constraint on who can actually use this is no longer legal, it is the hardware bill — and that is a very different bottleneck to argue about.