DeepSeek V4 drops 1.6T flagship, 1M context

DeepSeek V4 cements the trend of open-weights models matching or exceeding closed-source giants in raw scale and reasoning.

• –DeepSeek-V4-Pro's 1.6T MoE architecture (49B active) achieves parity with GPT-5.4 on MMLU-Pro, proving open-source can maintain frontier-level performance.
• –Standardizing a 1 million token context window across the family makes long-horizon agentic tasks a commodity for developers.
• –Hybrid Attention Architecture (CSA/HCA) significantly lowers the hardware cost of long-context inference by drastically reducing KV cache memory requirements.
• –Native optimization for Huawei Ascend 950 infrastructure highlights a strategic shift toward domestic hardware independence and localized compute.
• –MIT licensing of frontier-scale weights continues to pressure the pricing models and accessibility of proprietary ecosystems like OpenAI and Google.