AI Potluck
Model components / Base / pretrained models

DeepSeek-V4-Pro

DeepSeek

DeepSeek-V4-Pro, 1.6T total / 49B active MoE, 1M-token context, hybrid CSA+HCA attention. Released April 24 2026 under MIT, alongside V4-Flash (284B). Pretrained on more than 32T tokens.

V4-Pro-Max is the max-reasoning mode of V4-Pro rather than a separate release. Consolidated on 2026-07-29 from deepseek-v3-2, deepseek-v3-base, deepseek-v4-pro-base; openness follows deepseek-v4-pro, the release that currently governs. See docs/reference/identity.md. Adoption sums six declared checkpoints and V4-Flash outpulls V4-Pro, so the family reading is not the flagship reading. Verified 2026-08-13 via the DeepSeek-V4-Pro HF model card and the deepseek-ai/DeepSeek-V3 repository.

Openness

3 high confidence
3.0
weights
open(MIT, on HF, both Pro and Flash)
data
closed
code
partial(inference/serving
license
MIT(permissive, no use restrictions)

MIT weights across the V4 line, no released training corpus, and inference and serving code only, so 3. DeepSeek-V4-Pro is the current release, and both it and V4-Pro-Base are plain MIT with no acceptable-use restrictions attached. That keeps the score clear of an unsettled question: the older DeepSeek Model License carried acceptable-use terms, and whether those count as use-restricting is still open.

  • https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Pro recorded 2026-08-13

    card license `mit`, `gated: false`, `private: false`, safetensors weight shards listed, 1,385,596 downloads and 5,437 likes.

  • https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro recorded 2026-08-13

    Model card - "This repository and the model weights are licensed under the MIT License"; "We pre-train both models on more than 32T diverse and high-quality tokens", which describes the corpus without releasing it.

  • https://api.github.com/repos/deepseek-ai/DeepSeek-V3 recorded 2026-08-13

    `private: false`, `archived: false`, license `MIT`, 104,190 stars. Inference and serving code, no pretraining pipeline, which is what holds `code` at partial.

  • https://api-docs.deepseek.com/news/news260424 recorded 2026-08-15

    DeepSeek's own V4 release note, describing the line as shipping with 'Open Weights'. It announces weights and serving surfaces and releases no training corpus and no training pipeline, which is what holds the score at open_weights rather than higher.

  • https://simonwillison.net/2026/apr/24/deepseek-v4/ recorded 2026-08-15

    Third-party write-up of the V4 release recording 'MIT license' over 'open weights'. Corroborates the licence reading; establishes no dimension on its own, being a commentary rather than the artifact.

Adoption

4 high confidence
4.0

5,180,104 downloads in the trailing 30 days across the six declared artifacts (deepseek-ai/DeepSeek-V3.2 1,093,344; deepseek-ai/DeepSeek-V3-Base 31,718; deepseek-ai/DeepSeek-V4-Pro 1,411,583; deepseek-ai/DeepSeek-V4-Flash 2,283,943; deepseek-ai/DeepSeek-V4-Pro-Base 14,769; deepseek-ai/DeepSeek-V4-Flash-Base 344,747), which bands at level 4 (1M-10M) on the model adoption scale.

Capability

5 high confidence
5.0

SWE-bench Verified 80.6% (trails Claude Opus 4.6 by ~0.2), MMLU-Pro 87.5, GPQA Diamond 90.1, LiveCodeBench Pass@1 93.5 (reported best of any model). V4-Pro-Max described as the strongest open source model available, at/near closed frontier on coding and reasoning.

Verified 2026-08-13