DeepSeek-V4-Flash (284B total, 13B activated) and Pro (1.6T total, 49B activated) support 1M context length. The architecture uses DSA2 sparse attention, reducing inference FLOPs to just 10% and KV cache to 7% compared to V3.2.
The latest update (June 27) introduces DSpark, a speculative decoding framework that boosts inference speed by 60-85% at the same throughput, without changing output quality. The full-stack DeepSpec toolset is now open-sourced.
Top comments (2)
Want to access these models through one unified API? ThreeRouter gives you instant, enterprise-grade access to GLM-5.2, Qwen3.7, DeepSeek-V4, and more — all at a fraction of the cost. US-local deployment available. 1M free tokens for new users.
niu