DeepSeek V4.1 Flash Opens a Two-Day Test
DeepSeek V4.1 Flash entered an intermediate-version test on the afternoon of September 8. Its model name includes “expires-on-0910,” indicating that access is scheduled to end on September 10. This is not a general release, but a short evaluation window for an unfinished version. Developers can keep the existing base_url and change only the model name.
New Architecture and Native Multimodality
The central claim is not a routine parameter update but a new model structure. Unlike V4 Flash, which used a separate Vision extension for visual input, V4.1 Flash is described as natively multimodal, integrating image and text handling into the model itself. That could simplify the processing path, but real tasks are still needed to assess its capability, stability, and compatibility.
Pricing remains the same as V4 Flash, with no beta surcharge. During off-peak periods, cached input costs 0.05 yuan per million tokens, uncached input costs 1.5 yuan, and output costs 4.5 yuan. Peak-period prices are 0.1, 3, and 9 yuan respectively. Developers have reported speeds as high as 507 tokens per second, while an SVG animation example reached 328 tokens per second; several users reported averages above 300. These figures come from individual tests or community feedback and are not a universal guarantee.
The 20-Concurrency Cap Defines the Test’s Scope
The test version allows up to 20 concurrent requests per account, compared with 2,500 for the V4 Flash release and 500 for V4 Pro. That gap strongly suggests a purpose centered on feature checks, comparative evaluation, and feedback rather than production load testing or large-scale traffic. Developers should also examine time to first token, long-task stability, multimodal accuracy, and behavior across agent frameworks.
Can V4.1 Flash Take Over Pro Workloads?
The commercial question is straightforward: can the model approach Pro-level capability while retaining Flash-level cost and speed? The official feedback form specifically asks whether it could fully replace V4 Pro in production and requests account details, agent tools or frameworks, and comparative task results. If later versions perform well enough on complex reasoning and native multimodality, consumer applications and agent products could gain a lower-cost option. The available material, however, establishes an evaluation goal rather than proof of replacement.
DeepSeek made several V4-related moves from late July through August, including V4 Flash, V4 Pro, Harness, and Vision updates, so V4.1 Flash may mark the start of another iteration cycle. The cautious conclusion is that it is worth testing on short tasks and multimodal workflows before the deadline, but the evidence does not justify moving production traffic yet. Its lasting value will depend on whether a final release preserves the reported speed, pricing, and architectural advantages while removing or raising the concurrency limit.