AI Potluck
Contribute
Back to Gap Map Model components / Base / pretrained models

DeepSeek-V4-Pro

DeepSeek
open weights / Overall score: 4.7(leading)

DeepSeek-V4-Pro is DeepSeek's 1.6T-parameter mixture-of-experts model, with 49B parameters active per token, a one-million-token context and a hybrid attention design that combines Compressed Sparse Attention with Heavily Compressed Attention. It was pretrained on more than 32T tokens and ships as base and instruct checkpoints, alongside the smaller V4-Flash and the multimodal V4.1-Flash.

Openness follows the V4-Pro line, while capability is read on the family's highest-scoring releases and adoption across the V3 and V4 checkpoints this entry tracks.

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)

DeepSeek releases the V4 line's weights under a plain MIT license, but publishes no training corpus and only inference and serving code rather than a full training pipeline. This makes DeepSeek an open-weights release rather than a fully open model.

  • https://api-docs.deepseek.com/news/news260424 recorded 2026-09-29

    DeepSeek's V4 preview release note, which says "DeepSeek-V4 Preview is officially live & open-sourced!" and links the weights as "Open Weights" on the Hub collection. Searched for a training-data or training-code release: none, only weights, API and a tech report.

  • 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://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Pro recorded 2026-09-29

    Hub API record for deepseek-ai/DeepSeek-V4-Pro with gated false and private false, the tag "license:mit", and weight shards listed from "model-00001-of-00064.safetensors" onward. Searched the record for gated manual or an access prompt: none.

  • https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Pro-0813 recorded 2026-09-29

    Hub API record for deepseek-ai/DeepSeek-V4-Pro-0813, the official V4-Pro release, with gated false and private false, the tag "license:mit", and shards from "model-00001-of-00066.safetensors" onward.

  • https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro recorded 2026-09-29

    Model card states "This repository and the model weights are licensed under the" MIT License, and describes the corpus without releasing it: "diverse and high-quality tokens, followed by a comprehensive post-training pipeline", after "We pre-train both models on more than" 32T.

  • https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813/raw/main/README.md recorded 2026-09-29

    The V4-Pro-0813 card calls it "the official release of" V4-Pro, "superseding the preview version, with greatly enhanced agentic capabilities", and repeats "This repository and the model weights are licensed under the [MIT License](LICENSE)."

  • https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/inference/README.md recorded 2026-09-29

    The inference folder in the model repo is titled "Inference code for DeepSeek models" and covers weight conversion and generation ("First convert huggingface model weight files to the format of this project"). Searched for a pretraining script, data preparation or training recipe: none.

  • https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/LICENSE recorded 2026-09-29

    The LICENSE file shipped beside the weights is the plain MIT text, "Permission is hereby granted, free of charge, to any person obtaining a copy" with no use, field or scale restriction added.

  • https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/raw/main/README.md recorded 2026-09-29

    The V4.1-Flash card, the family's newest release, says it "is trained from scratch on a multimodal corpus comprising" 45T tokens without publishing that corpus, and that "This repository and the model weights are licensed under the [MIT License](LICENSE)."

  • https://simonwillison.net/2026/apr/24/deepseek-v4/ recorded 2026-09-29

    Third-party write-up of the V4 release: "They’re using the standard MIT license." Corroborates the license reading and establishes no dimension on its own.

Adoption

4 high confidence
4.0

Adoption is measured as Hugging Face downloads summed across the family's six shipped checkpoints, spanning the V3 line and the newer V4-Pro and V4-Flash releases. V4-Flash draws the largest share of that traffic, with the base checkpoints seeing comparatively little use.

Capability

5 high confidence
5.0

DeepSeek-V4.1-Flash scores 39 on Artificial Analysis's Intelligence Index and the official V4-Pro-0813 release scores 36, both among the highest-scoring large open-weight models on the index, though a few open-weight releases such as GLM-5.3 score higher. The V4-Pro preview card reports 80.6 on SWE-bench Verified and 93.5 on LiveCodeBench.

  • https://artificialanalysis.ai/models/deepseek-v4-1-flash recorded 2026-09-29

    Artificial Analysis page for the newest family member: "DeepSeek V4.1 Flash (Reasoning, Max Effort) scores 39 on the Artificial Analysis Intelligence Index", ranked 7th of the "116 models in this class", which for an open-weights model the page defines as "compared only with other open weights models of the same size class" (Large, over 150B).

  • https://artificialanalysis.ai/models/deepseek-v4-pro recorded 2026-09-29

    Artificial Analysis page for the official V4-Pro release: "DeepSeek V4 Pro 0813 (Reasoning, Max Effort) scores 36 on the Artificial Analysis Intelligence Index", ranked 9th of the "116 models in this class", which for an open-weights model the page defines as "compared only with other open weights models of the same size class", and it "is amongst the leading models in intelligence".

  • https://benchlm.ai/models/deepseek-v4-pro-high recorded 2026-09-29

    BenchLM profile (the URL now redirects to the V4-Pro-0813 page) still carries the DeepSeek-V4 technical-report row "Software Engineering Benchmark Verified" at "80.6%", beside an overall "BenchLM score 64.0 out of 100" for the official release.

  • https://codersera.com/blog/deepseek-v4-pro-review-benchmarks-pricing-2026/ recorded 2026-09-29

    Third-party V4-Pro review listing "80.6% SWE-bench Verified" and a LiveCodeBench row of 93.5% against Claude Opus 4.7 and GPT-5.5; corroborates the preview figures.

  • https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/README.md recorded 2026-09-29

    The V4-Pro preview card's V4-Pro-Max comparison table, with rows "| SWE Verified (Resolved) | **80.8** | - | 80.6 | 80.2 | - | 80.6 |" and "| LiveCodeBench (Pass@1) | 88.8 | - | 91.7 | 89.6 | - | **93.5** |", the primary source for the recorded preview figures.

  • https://www.morphllm.com/deepseek-v4 recorded 2026-06-04

    Third-party DeepSeek V4 write-up; the host answers HTTP 429.

Verified 2026-09-27