DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash, released September 10, 2026, opened a new architecture family for the lab — and did it from the bottom, as the smallest model in the line rather than a flagship. DeepSeek described it as built for greater capability, faster inference and higher throughput, with the architecture meant to scale up to larger models still to come, and with native visual understanding rather than a vision encoder attached after the fact. On the lab's own harness it scored 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1 and 88.1 on CyberGym, all ahead of the larger V4-Pro-0813 from four weeks earlier, and 54.8 on AutomationBench. Terminal-Bench 3.0 at 30.0 and 4.0 at 31.2 were around two and a half times the V4-Pro-0813 figures on the same boards.

The visual side of the launch was the genuinely new part for DeepSeek, whose V3 and V4 lines had been text-only: the launch table reported 78.9 on Chartography and 89.6 on BabyVision, both run with tools, putting a DeepSeek model on multimodal boards for the first time. Reasoning was the softer spot — 36.8% on Humanity's Last Exam without tools, the one unstarred figure on a row where the other DeepSeek columns reported the text-only subset, while 63.9% with tools was the highest in the comparison. As with the July and August snapshots, the code-agent numbers came from DeepSeek's own unreleased harness and its figures for rival models did not always match those labs' published ones, so the chart reads best as a within-family comparison. The launch thread announced no weight release.