DeepSeek-V4-Flash by DeepSeek — 1.0M Context

284B parameter (13B activated) Mixture-of-Experts language model with 1M token context length. Features hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), Manifold-Constrained Hyper-Connections (mHC), and Muon optimizer for faster convergence. Supports three reasoning effort modes: Non-think, Think High, and Think Max. Pre-trained on 32T+ tokens with comprehensive post-training via GRPO and on-policy distillation.