Model components / Fine-tuned / chat models

DeepSeek-V4-Pro / V4-Flash

DeepSeek's open-weight April 2026 flagship line, replacing V3.2. Released 2026-04-24 with 1M default context and 384K max output. V4-Pro-Max posts 80.6% on SWE-bench Verified (top-10) and a later snapshot reports 93.5% on LiveCodeBench. Available via the `deepseek-chat` (V4-Flash) and `deepseek-reasoner` aliases from 2026-07-24, with weights on Hugging Face under MIT. Demonstrates that open weights can match closed-API frontier coding performance at the same context window.

Compound SKU. DeepSeek-V4-Pro (1.6T total / 49B active MoE) is a FROZEN flagship anchor (base_pretrained shelf) - NOT rescored here; calibrate to it. This record scores the non-anchor sibling DeepSeek-V4-Flash (284B total / 13B active MoE, 1M context, FP4+FP8, hybrid CSA+HCA attention, mHC, Muon optimizer), the lighter instruct/thinking SKU of the V4 family. Verified live on HF June 2026.

Openness

3 high confidence

3.0

weights: open(MIT, on HF)
data: closed
code: partial(inference/serving
license: MIT(OSI-style permissive, no use restrictions)

V4-Flash weights are MIT open (matching the V4-Pro anchor's open_weights class); training data/code closed -> open_weights. Consistent with the frozen V4-Pro anchor's openness ladder.

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash recorded 2026-06-04
MIT license, 284B-A13B MoE, 1M context, downloadable FP4/FP8 weights

Adoption

4 high confidence

4.0

~3.50M downloads last month for V4-Flash on HF; the cheaper/faster sibling of V4-Pro is attractive for self-hosting, driving strong pull volume. Squarely 1-10M on measured monthly downloads (V4-Pro anchor itself sits at level 4).

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash recorded 2026-06-04
~3,503,796 downloads last month

Capability

4 medium confidence

4.0

Near-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 2026-06-04
MMLU-Pro 86.2, LiveCodeBench 91.6, MRCR-1M 78.7 (Max mode); base MMLU 88.7

Unchanged since 2026-07-29 (last edited, not re-checked)