DeepSeek V4 GA: 1.6T params, FP4 inference, API at $0.14/M tokens

DeepSeek V4 GA is out: MoE with 1.6T params (49B active), hybrid attention CSA+HCA, FP4 Lightning Indexer reducing KV cache by 98%, 1M context. Coding benchmarks leading, API $0.14/M tokens (14-29x cheaper than Claude). Open source under MIT.

• DeepSeek V4
• 开源
• MoE
• CSA
• HCA
• FP4