DeepSeek V4.1 Flash has now been updated to version V4.1-Flash-1017, and the official web interface, app, and API models have all been switched to the new model in sync.
A Pareto Improvement in Performance vs. Cost
Compared with the previous version of DeepSeek V4.1 Flash, the new model has made substantial gains across benchmarks in every area.
While improving performance, we also shortened the model's task runtime and output length, letting users complete tasks at lower cost and higher speed (see the figure below), reaching a new performance–cost Pareto optimum under multiple harnesses. In particular, this time we focused our optimization efforts on platform frameworks such as DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. Please upgrade to DeepSeek Harness 0.2.1-rc.1 or above to get the best experience with the new model.
Performance/cost chart for DeepSeek V4.1 Flash 1017 running DeepSWE on the mini-swe-v2 framework.
Complete Project Development from Scratch
Through reinforcement learning (RL) training based on self-feedback, DeepSeek V4.1 Flash 1017 has improved its ability to build complete projects from scratch in areas such as game development and 3D modeling, and can support more "wish-style" programming tasks and project development requests. The following cases show what the new model achieves inside the DeepSeek Harness framework with very few prompts and minimal reference materials, in a very small number of interaction rounds:
Example 1: 3D modeling showcase of the Forbidden City
Example 2: Web-based racing game developed by the model
Example 3: A recreation of Portal 2 built on the Godot engine
Example 4: Valley of Ages — a dynamic 3D scene tracing the arc of development from prehistoric times to the modern era
Opening Up Our In-House Search API
For a long time, the search capabilities on DeepSeek's web and API sides have been powered at the underlying level by our in-house DeepSeek search engine, which reliably handles hundreds of millions of search requests from around the world every day.
DeepSeek Search is an AI-native general-purpose search engine that indexes hundreds of billions of web pages worldwide, supports multilingual search, and returns high-relevance, high-authority, and high-timeliness quality results in real time. Against SEO cheating, GEO manipulation, and low-quality AIGC, we apply strict identification mechanisms to filter them out, ensuring that authoritative and trustworthy information is presented to users and to LLMs.
This new release adds support for the Search and Fetch interfaces to the DeepSeek API, formally opening these capabilities to the public. Both interfaces are priced at 5 yuan per thousand calls. For detailed usage instructions, please refer to the official API documentation: API Guide – Web Search and Fetch.
Please note that users currently accessing the DeepSeek API within DeepSeek Harness / Claude Code / Codex already reach the Search/Fetch interfaces through server-side tool calls. These Search/Fetch interface calls are currently not billed, and starting in two weeks (00:00 Beijing time, November 1, 2026) they will be charged per call.
Open-Source Model
The download addresses for the DeepSeek-V4.1-Flash-1017 model weights are as follows:
HuggingFace:
https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
ModelScope:
https://modelscope.cn/models/deepseek-ai/DeepSeek-V4.1-Flash-1017