Tier 3 Practical open-weight
What the publisher claims: open-source, open weights, open models
The one-liner: Permissive and influential open-weight family with substantial technical documentation and unresolved provenance debates.
The six questions
Weights for V4 Pro, V4 Flash, and V4.1 Flash are downloadable from Hugging Face and ModelScope, along with the inference code.
MIT on the weights and the accompanying code, with no use policy, attribution, or revenue condition attached to the current releases.
Independently deployable, with published inference engines, FP8 builds, and deployment guidance; the 284B Flash checkpoint reaches a single B200-class node.
Technical reports disclose architecture (hybrid compressed and hybrid cross attention, manifold-constrained hyper-connections, Muon at 1.6T scale), pretraining token counts of 32-33T, training stages, and DeepSeek publishes a formal model-mechanism disclosure with a V4 training-data summary; the corpus itself is not published.
Moderate for research replication: the recipe, token counts, and architecture are public and the kernels are designed for bitwise-reproducible pipelines, but the training data is not.
deepseek-recipe and the Hugging Face model repos accept issue reports and the occasional outside pull request, but the model roadmap, training data, and release train are set entirely by DeepSeek.
License
MIT for the model weights and the accompanying code (deepseek-recipe, DeepInfer) OSI-approved — license source ↗
- Earlier V3/R1-era repositories mix MIT code with a custom model license; verify each derivative
What is actually released
- weights (V4 Pro, V4 Flash, V4.1 Flash and the 1017 update)
- inference code (deepseek-recipe, DeepInfer)
- technical reports with architecture and token counts
- formal training-data summary and model-mechanism disclosure
- evaluation results
Why it matters
DeepSeek combines strong practical openness, permissive terms, efficient training claims, and controversy over possible distillation - capability, license, provenance, and allegations must stay separate.
Sources
- https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
- https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- https://arxiv.org/html/2606.19348v1
- https://cdn.deepseek.com/policies/en-US/model-algorithm-disclosure.html
- https://github.com/deepseek-ai/deepseek-recipe
- https://lambda.ai/blog/deepseek-v4-the-most-expected-open-source-model
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.