DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, describing it in its API change log as “the smallest model in our new architecture family, with native multimodal visual understanding.” The release also retired the previous V4 Flash and V4 Flash Vision Exp models.
On this page
What changes in the API
DeepSeek says developers reach the latest V4.1 Flash with the model name deepseek-flash. The earlier names deepseek-v4-flash and deepseek-v4-flash-vision-exp are “temporarily routed to V4.1 Flash,” so existing integrations keep working while they move to the new name. Because the routing is described as temporary, applications that still call the old names have a reason to move to deepseek-flash.
The same entry notes that DeepSeek V4 Pro API services continue beyond September 14, 2026.
Benchmark results reported by DeepSeek
DeepSeek published a set of its own scores for V4.1-Flash alongside the release. A selection:
| Benchmark | DeepSeek-V4.1-Flash |
|---|---|
| GPQA Diamond | 90.9 |
| HLE | 36.8 |
| HLE (with tools) | 63.9 |
| Terminal-Bench 2.1 | 90.6 |
| Terminal-Bench 4.0 | 31.2 |
| DeepSWE v1.1 | 74.2 |
| CyberGym | 88.1 |
| Codeforces (rating) | 3471 |
All figures are DeepSeek’s own. The list also includes vision tasks run with tools, such as BabyVision (89.6) and Chartography (78.9), which reflect the model’s new image input.
Context: V4 Pro reached general availability in August
The Flash release follows the general availability of DeepSeek-V4-Pro, which DeepSeek announced on August 13, 2026. That release rolled out across DeepSeek’s app, web and API, with calls still made through deepseek-v4-pro. DeepSeek said the GA version “greatly enhances agent capabilities,” with the largest gains in production environments.
The August update brought two platform changes that also apply to the Flash line. The API “natively supports the OpenAI Responses API format,” adapted for Codex, and V4-Pro and V4-Flash gained “three thinking effort levels: low / high / max,” which developers choose per task.
What it changes
For developers on the Flash tier, the September 10 release means a model swap behind an unchanged integration: requests to the old V4 Flash names now reach V4.1 Flash, and image understanding is native to the model instead of a separate experimental variant. Teams that pinned deepseek-v4-flash-vision-exp for vision work are now served by the same model as text requests.
The release is listed in our AI model release timeline, and the current state of DeepSeek’s service is on the DeepSeek status page.
Sources
- Change Log | DeepSeek API Docs — DeepSeek
