DeepSeek has officially introduced the V4.1-Flash-1017 model, which is a minor update to its V4.1 Flash model. Consequently, the company has switched its services for the web interface, mobile application, and API to the new version. API users do not need to make any changes, as the model continues to be referred to as deepseek-flash, and the pricing policy remains the same, according to the company's statement reported by IT Home.
The model weights are available on Hugging Face and ModelScope under the MIT license, consistent with the terms applicable to V4.1 Flash. Previously, on October 17th, the 'deepseek-v4.1-flash-1017' entry briefly appeared in the DeepSeek API model list, as reported by Pandaily, but this was before the official announcement. This launch now confirms that build. Bloomberg and The Straits Times reported on the update, noting its higher speed and reduced cost. Chinese sources provided additional information regarding the training process, model tuning, and context of its use.
DeepSeek characterizes this update as a Pareto improvement because it achieves increased benchmark scores while simultaneously reducing execution time and output length, thereby lowering the overall cost per call. According to the model card on Hugging Face, the score on the Terminal-Bench 4.0 test (using the Claude Code handler) increased from 31.2 to 42.3. Similarly, ProgramBench rose from 20.3 to 29.1, DeepSWE v1.1—from 74.2 to 77.0, NL2Repo-Bench—from 65.4 to 68.2, and Automation-Bench—from 54.8 to 60.5.
The key functional change, which DeepSeek calls wish-style programming, is the ability to create a complete, executable project from scratch using only a brief prompt and minimal reference materials. The company explains this through the application of self-reflection reinforcement learning, during which the model critiques its own final result during training. Demonstrations hosted on the DeepSeek website include a 3D model of the Museum Palace, a browser racing game, a recreation of Portal 2 in the Godot engine, and an epoch-spanning scene called Valley of Ages, all created in very few iterations within DeepSeek Harness.
Furthermore, DeepSeek has adapted the model specifically for DeepSeek Harness, as well as for Tencent's tools—CodeBuddy and WorkBuddy. For full support, users require version 0.2.1-rc.1 or newer of Harness. Alongside the model itself, the company has provided developers access to its internal search engine via separate API Search and Fetch endpoints.
This release followed service disruptions. On the morning of October 15th, users discovered that DeepSeek was unavailable for nearly five hours, according to a Tencent News report. According to DeepSeek's public status page, 44 incidents of varying severity have been registered this year, including brief performance degradation and partial outages, one of which lasted about 12 hours at the end of March. DeepSeek did not link this outage to the model update.