DeepSeek-V4-Pro / V4-Flash
DeepSeekDeepSeek's open-weight April 2026 flagship line, replacing the previous V3 generation, with a one-million-token default context and up to 384K output. It is reached through the deepseek-chat and deepseek-reasoner aliases as well as weights published on Hugging Face.
A compound record. The V4-Pro flagship is a frozen anchor on the base_pretrained shelf and is not rescored here; this record scores the lighter V4-Flash sibling, a 284B mixture-of-experts with 13B active parameters. Verified 2026-08-13 via the DeepSeek-V4-Flash model card.
Openness
3 high confidence- weights
- open(MIT, on HF)
- data
- closed
- code
- partial(inference/serving
- license
- MIT(OSI-style permissive, no use restrictions)
V4-Flash weights are MIT-licensed and downloadable, matching the V4-Pro flagship, while the training data and code stay closed - open weights rather than open source, on the same rung V4-Pro sits on.
- https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash recorded 2026-08-13
License section - "This repository and the model weights are licensed under the MIT License"; `license:mit` in the repo tags and `"gated":false` in the embedded repo state, so the 284B-A13B FP4+FP8 weights download without a barrier; encoding and local-inference folders only, with the post-training pipeline described but not released and no SFT/RL mixture published
Adoption
4 high confidenceV4-Flash reads 2,213,519 Hugging Face downloads in the trailing 30 days and the declared V4-Pro repository reads 1,385,596, so both sit inside the 1M-10M band and the level holds whichever SKU the tier is read on. Flash is the cheaper, faster sibling of V4-Pro and attractive for self-hosting, which is what drives the volume.
- https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash recorded 2026-08-13
2,213,519 downloads in the trailing 30 days
Capability
4 medium confidenceNear-frontier capability (MMLU-Pro 86.2, LiveCodeBench 91.6) but deliberately the lighter sibling of the V4-Pro anchor (capability 5); placed at 4 to keep the family curve consistent - Flash trails Pro. Figures DeepSeek-reported on the HF card.
- https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash recorded 2026-08-13
Comparison across Modes - V4-Flash Max MMLU-Pro 86.2, LiveCodeBench 91.6, MRCR 1M 78.7, SWE Verified 79.0; base-model table MMLU 88.7, HumanEval 69.5, GSM8K 90.8
Verified 2026-08-13