This market resolves YES if ALL of the following things are true of Deepseek V4 (and V4-lite):
V4 1.6T, V4-Lite 285B
Attention: DSA2 (NSA + DSA), head-dim 512 Sparse MQA + SWA
MoE: Fused MoE Mega-Kernel with 6 active in 384 experts
Residual: Hyper-Connections
Optimizer: Muon
Pretrain context length: 32K
RL: GRPO with corrected KL
Final Context Length: 1M
Modality: Text only

Statements that cannot be confirmed after the release of the paper/model will be ignored (for example, we dont learn about the pretrain CL) and will not count for a YES/NO resolution. If there are disagreements about whether an item is true or not that are not solved with a reasonable amount of discussion, i will feed V4 itself the V4 paper along this market criteria and use its response as the final source of truth.

• Update 2026-04-23 (PST) (AI summary of creator comment): The creator is leaning NO due to V4-Flash being 284B rather than the predicted 285B. Resolution expected in a few days.