DeepSeek V4.1 Flash preview: Pro-class quality at Flash prices
Callable now as a two-day API beta at ~420 tok/s with native image input; DeepSeek says V4-Pro traffic will route to it at ~3x cheaper output.
Copy markdown
Live now, gone by Sept 10
The preview is callable on the standard DeepSeek API as model `deepseek-v4.1-flash-expires-on-0910` — no waitlist, no new endpoint. The beta window closes around Sept 10, so benchmark it against your V4-Pro prompts today, not next week.
Pro-tier reasoning at Flash-tier cost
DeepSeek says V4.1 Flash “comprehensively surpassed V4 Pro in performance, cost, speed.” It bills at V4-Flash rates — roughly $0.14/1M in and $0.28/1M out versus Pro’s ~$0.44/$0.87 — so if the claim holds, that’s Pro-grade output for about a third of the price.
Fast, and now natively multimodal
Community runs clock ~420 tok/s with peaks past 500, and unlike V4-Flash’s bolt-on vision encoder, the preview takes image-and-text input natively. Context length and parameter count aren’t disclosed yet.
The routing change to watch
Right now nothing auto-switches — the beta is opt-in. But DeepSeek says that once V4.1 Flash launches it will route all V4-Pro requests to it at the Flash price until a V4.1 Pro ships. If you run V4-Pro in production, price and behavior are about to move under you.
What you’re comparing against
For reference, V4-Pro is a 1.6T-param MoE (49B active) at ~80.6% on SWE-bench Verified; V4-Flash is 284B/13B, self-reporting 73.7%. Both ship MIT open weights on Hugging Face with a 1M-token context — but V4.1 Flash weights aren’t out, so this preview is API-only for now.