DeepSeek has opened an intermediate preview of DeepSeek V4.1 Flash for internal testing. The official description says it uses a new model architecture with native multimodal support, delivering stronger capability, faster speed, and lower cost.
Access has almost no barrier to entry: the base_url stays the same, and calling the model just requires switching the model name to deepseek-v4.1-flash-expires-on-0910. Billing matches deepseek-v4-flash, with each account capped at 20 concurrent requests. The "0910" in the name marks the validity window — after September 10, this model ID stops working.
Native multimodality is the real point of this change
Previously, feeding mixed image-and-text input into the Flash series required a separate vision extension package. This version builds that capability directly into the model itself.
The difference is more than just dropping one extra package. The extension-package approach typically runs images through a separate encoder first, then splices the result into the text sequence — that conversion step costs both latency and accuracy. Native multimodality means images and text are trained and run inference within the same architecture. For a product line like Flash, built around low latency, this is a directional shift: the goal isn't a higher ceiling on visual understanding, it's not paying extra latency for multimodal support.
By positioning, V4.1 Flash sits between the existing V4 Flash and V4 Pro: an updated architecture with native multimodality, while keeping the low-latency trait the Flash series has always had.
Putting the expiration date inside the model name
This naming convention isn't new for DeepSeek, but applying it to an "intermediate version" carries more information than the announcement itself.
A model ID with a built-in expiration date means anyone integrating it must finish evaluation within two days and know it's coming out. That rules out any production deployment, leaving essentially one purpose: collecting performance data on this new architecture under real traffic. The 48-hour window and the 20-concurrent-per-account cap frame this as a stress test, not a release.
For developers in China, there's limited judgment to be made in practice. Parameter count, context-length changes, and evaluation results have not been disclosed officially, and there's no word yet on V4.1's official release timeline or whether this new architecture will eventually reach the Pro tier. "Faster and cheaper" is, for now, only the vendor's own claim — it can't be independently verified. The one thing that can be checked directly is that the price hasn't changed: same money, two days of runway.
It's worth comparing this to the product line's past pace. Moving V4-Flash from preview to a full API release, locking V4-Pro's promotional price into the permanent price list, and adding peak-hour double pricing to the API — all of these happened within a few months of each other. DeepSeek has never given much advance notice on version rollouts, and this time it hasn't even given a timeline for the official release.
Sources: DeepSeek Open Platform announcement, IT Home, CocoLoop, The Paper; the model ID, billing basis, and concurrency limit follow the official preview notes, while parameter count and evaluation data have not been disclosed officially.