DeepSeek V4 Pro is DeepSeek's flagship open-source frontier model, released April 24, 2026, with 1.6 trillion total parameters (49B active) — the largest open-weight model to date. It introduces a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), reducing single-token FLOPs to 27% and KV cache to 10% compared to V3.2 at 1M-token context. Trained with mixed FP4/FP8 precision and Manifold-Constrained Hyper-Connections, it supports dual Thinking and Non-Thinking modes across a 1M-token context window. At launch, it scored 80.6% on SWE-bench Verified (tying Claude Opus 4.6), 93.5% on LiveCodeBench, and a 3206 Codeforces rating, while costing roughly one-seventh the price of comparable frontier models.
DeepSeekDeepSeek V4 Pro is DeepSeek's flagship open-source frontier model, released April 24, 2026, with 1.6 trillion total parameters (49B active) — the largest open-weight model to date. It introduces a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), reducing single-token FLOPs to 27% and KV cache to 10% compared to V3.2 at 1M-token context. Trained with mixed FP4/FP8 precision and Manifold-Constrained Hyper-Connections, it supports dual Thinking and Non-Thinking modes across a 1M-token context window. At launch, it scored 80.6% on SWE-bench Verified (tying Claude Opus 4.6), 93.5% on LiveCodeBench, and a 3206 Codeforces rating, while costing roughly one-seventh the price of comparable frontier models.
DeepSeek