DeepSeek rolled out V4.1 Flash 1017 across its app, website and API, reporting improved coding, agent and vision benchmark scores.
DeepSeek announced V4.1 Flash 1017 on October 17, making the updated model available through its website, mobile app and API. The company also published downloadable model weights under the MIT license, enabling developers to run or adapt the release themselves. DeepSeek describes this as an update to V4.1 Flash, rather than a new model family.
The company’s model card reports gains on several benchmarks compared with the previous V4.1 Flash version. Its Terminal-Bench 4.0 score rose from 31.2 to 42.3, while DeepSWE v1.1 increased from 74.2 to 77.0. The card also lists improvements on visual and agent-oriented evaluations, including BabyVision with tools, which moved from 89.6 to 92.5, and Automation-Bench, which rose from 54.8 to 60.5. These are DeepSeek-reported results; the published figures do not, by themselves, establish independent verification. The card notes that Terminal-Bench used the Claude Code harness and DeepSWE used the Mini-SWE-V2 harness.
DeepSeek says the update is designed to complete tasks faster and with shorter outputs, potentially lowering time and token use in agent workflows. It highlights compatibility work for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and says the existing API model name, `deepseek-flash`, continues to serve the current version. The published model card specifies a context window of up to one million tokens.
The release may be relevant to developers comparing open-weight models for coding, tool use and multimodal tasks: it combines updated benchmark claims with public weights and availability through DeepSeek’s existing API. Practical results will still depend on the chosen harness, deployment setup and workload. DeepSeek’s announcement provides benchmark comparisons and examples of end-to-end project generation, but does not publish an independent evaluation of the new snapshot.