DeepSeek V4 Pro

DeepSeek-V4-Pro-DSpark is a text-generation model from the DeepSeek-V4 series, a Mixture-of-Experts (MoE) language model with 1.6 trillion total parameters (49B activated) supporting a context length of one million tokens. It incorporates a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to improve long-context efficiency, Manifold-Constrained Hyper-Connections (mHC) for enhanced signal propagation, and the Muon optimizer for training stability. The model was pre-trained on over 32 trillion tokens and post-trained via a two-stage paradigm of domain-specific expert cultivation followed by unified consolidation through on-policy distillation.

Key Strengths

In the 1M-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2. The model supports three reasoning effort modes: Non-think (fast, intuitive), Think High (conscious logical analysis), and Think Max (maximum reasoning).

Benchmark Results

DeepSeek-V4-Pro-Max achieves competitive results against frontier models across knowledge, reasoning, coding, and agentic benchmarks:

• LiveCodeBench (Pass@1): 93.5 (highest among compared models)
• Codeforces (Rating): 3206 (highest among compared models)
• Apex Shortlist (Pass@1): 90.2 (highest among compared models)
• MMLU-Pro (EM): 87.5
• GPQA Diamond (Pass@1): 90.1
• SWE Verified (Resolved): 80.6
• LongBench-V2 (EM): 51.5 (base model)

DeepSeek-V4-Pro-DSpark is the same checkpoint with an additional speculative decoding module attached for faster inference. It is hosted as an OpenAI-compatible API on gigarouter.