DeepSeek V4.1 Flash Opens a Short-Lived Limited Beta
A temporary DeepSeek V4.1 Flash route is available for limited testing through the existing API base URL. The reported beta keeps V4 Flash billing, caps each account at 20 concurrent requests, and uses a model ID that points to a September 10 expiry.
A DeepSeek V4.1 Flash intermediate build has entered limited testing with a temporary API model ID. Testers can keep their existing base_url and change only the model name to deepseek-v4.1-flash-expires-on-0910.
The access details surfaced publicly on September 8 in community posts reproducing a chat-group notice. DeepSeek has not yet attached the beta to a public release note or model documentation page, so this should be treated as a short-lived test route rather than a generally available model launch.
The only required API change is the model name
Existing DeepSeek API users do not need a new endpoint for the beta. Keep the current base URL, authentication, and request flow, then send the temporary model ID in the request body:
{
"model": "deepseek-v4.1-flash-expires-on-0910"
}
The model name itself signals an expected September 10 expiry. No public notice currently specifies the exact cutoff time, timezone, or whether the route could be extended. Applications should therefore keep the model ID configurable and retain deepseek-v4-flash as the rollback path instead of hard-coding this beta into production.
A new architecture with native multimodal support
The test notice describes V4.1 Flash as an intermediate version built on a new model architecture. It claims native multimodal support, stronger capability, faster responses, and lower cost.
Those points describe the direction of the beta, not a completed public specification. DeepSeek has not published a V4.1 Flash model card, benchmark table, context-window specification, supported media list, or multimodal request schema for this temporary route. Developers should verify text quality, latency, tool use, and any image-input behavior with their own test cases before drawing comparisons with the current V4 Flash release.
The lower-cost claim also does not mean testers receive a lower API price today. Billing for the temporary route is currently the same as deepseek-v4-flash. The new architecture may reduce serving cost, but no separate V4.1 Flash customer price has been announced.
Each account is limited to 20 concurrent requests
The beta caps each account at 20 concurrent requests. That limit is enough for interactive evaluation and small parallel test suites, but it is not a production throughput commitment. Teams running automated evaluations should bound concurrency, log rate-limit responses, and compare results against the existing Flash route under the same prompts and sampling settings.
A useful evaluation should separate four questions: whether the model follows instructions more reliably, whether native multimodal requests work consistently, whether time to first token and generation speed improve, and whether the same task costs less at the currently unchanged API price. Without that separation, a faster response can be mistaken for a capability gain, or lower infrastructure cost can be mistaken for a customer price cut.
This is a beta signal, not general availability
The temporary model ID and two-day window make the operating boundary unusually clear. DeepSeek V4.1 Flash is available to try, but the beta does not establish a permanent model name, stable multimodal contract, final price, production rate limit, or public release date.
Developers who test it should save prompts, outputs, latency measurements, token usage, error responses, and the exact test time. If the route disappears on September 10 as its name suggests, those records will be the only reliable way to compare the experiment with a later V4.1 Flash release.