DeepSeek-V4-Pro / V4-Flash

weights

open(MIT, on HF)

data

closed

code

partial(inference/serving

license

MIT(OSI-style permissive, no use restrictions)

V4-Flash ships MIT-licensed weights, matching the V4-Pro flagship, but the training data stays closed and only inference and serving code is public. That keeps it an open-weights release rather than open source, the same class as V4-Pro.

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

  Hugging Face model API: "gated":false and the tag "license:mit"; repository files include "inference/model.py" and "inference/generate.py" and "encoding/encoding_dsv4.py" but no training code, data or dataset files.

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

  README License section: "This repository and the model weights are licensed under the [MIT License](LICENSE)."; front matter "license: mit". Searched the body for training data release, dataset and training code: none; it describes the post-training pipeline and points only to the "encoding" and "inference" folders for running the model.

Adoption sums Hugging Face downloads across the two declared checkpoints, V4-Pro and V4-Flash, as the family rule requires; V4-Flash draws roughly twice V4-Pro's downloads.

• https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Flash recorded 2026-10-08

  "downloads":966499 over the trailing 30 days for deepseek-ai/DeepSeek-V4-Flash.

• https://huggingface.co/api/models/deepseek-ai/DeepSeek-V4-Pro recorded 2026-10-08

  "downloads":423640 over the trailing 30 days for deepseek-ai/DeepSeek-V4-Pro.

V4-Flash scores near the frontier on DeepSeek's own benchmarks (MMLU-Pro 86.2, LiveCodeBench 91.6), figures reported on its Hugging Face model card. It is deliberately the lighter, faster sibling of the higher-scoring V4-Pro flagship, so Flash trails Pro rather than matching it.

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

  Model table lists "DeepSeek-V4-Flash | 284B | 13B | 1M | FP4 + FP8 Mixed"; instruct comparison row "MMLU-Pro (EM) | 83.0 | 86.4 | 86.2"; "LiveCodeBench (Pass@1) | 55.2 | 88.4 | 91.6"; "MRCR 1M (MMR) | 37.5 | 76.9 | 78.7"; "SWE Verified (Resolved) | 73.7 | 78.6 | 79.0"; base table "HumanEval (Pass@1) | 0-shot | 62.8 | 69.5 | **76.8**".