DeepSeek-V4

Made by AIWritten by AI (Claude) from the source below. Check anything important against the source.How it's made

DeepSeek-V4 — CSA and HCA hybrid attention, mHC, Muon, 32T+ tokens

Why it matters

At 27% of the per-token FLOPs and 10% of the KV cache of V3.2, it made 1M context something supported all the time

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2606.19348

Added in Week 35, 2026 · See that week

On the map — ⑪ Model Architecture line

  1. Previous stopMistral Small 4
  2. This recordDeepSeek-V4
  3. Next stopIntern-NCP Team
See it on the map

Same week

DeepSeek releases V4-Pro (1.6T/49B) and V4-Flash (284B/13B), MIT, 1M contextOfficialCohere announces acquisition of and merger with Germany's Aleph AlphaPressGPT-5.5OfficialGoogle Cloud previews its 8th-generation TPUs at Cloud NextPress

Nearby on the same axis

Mistral AI brings in physics AI by acquiring and integrating Emmi AIOfficialMistral AI releases Mistral Small 4OfficialSkalingPaperNested Learning / HOPEPaper