DeepSeek V4 packs 1.6T parameters, 1M context

DeepSeek V4 is a "price-performance" hammer aimed directly at GPT-5 and Gemini 3.1, proving that architectural efficiency can rival raw compute scale.

• –1M token context window uses "Compressed Sparse Attention" to keep KV cache memory requirements 90% lower than standard transformers
• –V4-Pro matches GPT-5.4 in coding benchmarks while being significantly cheaper to run via API or self-hosting
• –Optimization for Huawei Ascend 950 silicon signals a maturing domestic Chinese hardware ecosystem independent of NVIDIA export bans
• –MIT-licensed weights for the 284B Flash model set a new high bar for local long-document processing and RAG applications
• –"Friday afternoon surprise" release strategy continues to disrupt Western AI news cycles with open-weight alternatives