DeepSeek V4 tops benchmarks, rivals GPT-5
DeepSeek V4 is a flagship 1.6-trillion parameter Mixture-of-Experts (MoE) model that sets new open-source records, matching or exceeding GPT-5 and Claude 4.5 in coding and mathematical reasoning. Featuring a native 1-million token context window and a novel "Engram" memory architecture, it delivers state-of-the-art performance with significantly higher inference efficiency than its closed-source peers.
DeepSeek V4 effectively ends the era of closed-source dominance in reasoning and coding tasks. Achieving 81% on SWE-bench Verified positions it as the first open-source model to consistently beat proprietary leaders like Claude 4.5. The Engram memory architecture enables a massive 1M token window with 97% retrieval accuracy, solving the "lost in the middle" problem, while Deeply Sparse Attention (DSA) keeps inference costs flat as context scales. Native multimodality pre-training ensures more coherent cross-modal reasoning than previous patched solutions.
DISCOVERED
5h ago
2026-04-24
PUBLISHED
6h ago
2026-04-24
RELEVANCE
AUTHOR
bigboyparpa