Btw I think V4-Pro has modestly accelerated V4-Flash is at 100 t/s, and indeed seems more token-effective This is really nice after what feels like years of DS API at 22 t/s
4:13 PM · Jun 16, 2026 · 2.3K Views
DeepSeek V4-Flash API reportedly reaches speeds of 100 tokens per second, up from a 22 t/s baseline
Engineers are debating the underlying optimizations driving the speed gains
14119099.5K
Sentiment
Sentiment building, check back later.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Posts from X
Most Activity
Most Activity
VIEWS4.5KBOOKMARKS5LIKES54REPLIES6
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
Why would DeepSeek get 22-40% faster? saw up to 110 t/s Flash, up to 90+ on Pro Inference optimizations, like at other labs? I would think they've already optimized the hell of it for RL alone, they had built this architecture for speed. Smaller bs? Or new hardware at last?
5hViews 4.5KLikes 54Bookmarks 5
kalomaze@kalomaze
@teortaxesTex better speculation?
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
Why would DeepSeek get 22-40% faster? saw up to 110 t/s Flash, up to 90+ on Pro Inference optimizations, like at other labs? I would think they've already optimized the hell of it for RL alone, they had built this architecture for speed. Smaller bs? Or new hardware at last?
5hViews 950Likes 14Bookmarks 0
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
…on second thought I guess it might be optimization actually
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
Why would DeepSeek get 22-40% faster? saw up to 110 t/s Flash, up to 90+ on Pro Inference optimizations, like at other labs? I would think they've already optimized the hell of it for RL alone, they had built this architecture for speed. Smaller bs? Or new hardware at last?
5hViews 1.8KLikes 15Bookmarks 0
Binkg@dummerspast39
@teortaxesTex will it be today?
7hViews 68
Odsii@who_odsi
@teortaxesTex ds api was 30ts-40ts for me
7hViews 41