1.6T-param MoE. The reference point for what frontier capability should cost.
DeepSeek V4-Pro is the strongest open-weight model in the survey and one of its most consequential entries. It is MIT-licensed — genuinely downloadable, self-hostable, and fine-tunable — and posts an independent intelligence index of 53, the best independently verified mark of any open model here. At $0.44 / $0.87 per million tokens on the hosted API, it functions as a reference point for what frontier-adjacent capability should cost.
The trade-offs are real but narrow. It is text-only, with no native image, audio, or video; the hosted API carries a peak-hours surcharge; and the 1.6T-parameter mixture-of-experts is heavy to self-host at full size. On the very hardest tasks it sits below the closed flagships.
For teams that need open weights — for control, privacy, on-prem deployment, or cost — V4-Pro is the current benchmark, and the 1M-token context and the free DeepSeek app that runs on it underline how much capability is now available without a closed vendor.
FIELD NOTE — Text-focused; peak-hours surcharge. Powers the free DeepSeek app.
VISIT OFFICIAL SITE →