DeepSeek V4 Flash DSpark

DeepSeek-V4-Flash-DSpark is a text-generation model, specifically a 284B-parameter Mixture-of-Experts (MoE) language model with 13B activated parameters that supports a context length of one million tokens and includes an attached DSpark speculative decoding module for improved inference throughput. The architecture combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) in a hybrid mechanism, together with Manifold-Constrained Hyper-Connections (mHC) and the Muon optimizer. In the 1M-token context setting, this model requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2. It was pre-trained on more than 32T tokens. On base model benchmarks, the model achieves MMLU 88.7 (5-shot), HumanEval 69.5 (0-shot Pass@1), GSM8K 90.8 (8-shot), MATH 57.4 (4-shot), and LongBench-V2 44.7 (1-shot). The instruct model supports three reasoning effort modes: Non-Think, Think High, and Think Max. In Max mode, it achieves LiveCodeBench 91.6 (Pass@1), GPQA Diamond 88.1 (Pass@1), HMMT 2026 Feb 94.8 (Pass@1), and SWE Verified 79.0% resolved. In Non-Think mode, it scores MMLU-Pro 83.0 and GPQA Diamond 71.2. The model supports tool calling across all modes.