The Frontier · Tags

#long-context

2 published articles

• Models · Ground Truth

  One million tokens fixes long-document recall

  DeepSeek-V4 reports MRCR 1M MMR 83.5 and CorpusQA 1M ACC 62.0 at Pro Max. Strong on paper. Not a buyer corpus guarantee.

• Compute & power · The Record

  DeepSeek-V4 paper reports 27% FLOPs and 10% KV cache at 1M tokens versus V3.2

  arXiv:2606.19348 (April 26, 2026); MIT checkpoints on HuggingFace; CSA/HCA hybrid attention.