DeepSeek released V4.1 Flash on September 10, 2026, adding native image understanding to its general-purpose API model. The current API name is deepseek-flash. Calls made using the older deepseek-v4-flash and experimental vision model IDs temporarily route to the new version.
V4.1 Flash uses a mixture-of-experts architecture with 552 billion parameters in total, activating about 8 billion for input and 16 billion for output. DeepSeek reports its key-value cache requires one-quarter of the previous generation's high-bandwidth memory and one-eighth of its SSD storage.
For coding-agent builders, DeepSeek reports a 90.6 score on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1 in its published results. The numbers describe its benchmark setups, not performance guarantees for a particular repository or agent.
API pricing also changed on September 10, with off-peak rates set at half the peak rates. DeepSeek's initial announcement said V4 Pro calls would eventually route to Flash, but the API changelog later clarified that V4 Pro service would continue after September 14 with unchanged billing. Developers depending on that model should follow the updated API notice.