文章摘要
DeepSeek推出升级大模型V4.1-Flash-1017，参数规模7630亿，多项基准测试性能提升，权重以MIT协议开源，支持百万token上下文与图文混输。该模型任务成本远低于旧版及竞品，支持提示词驱动从零搭建各类项目；此次同步全面开放自研搜索API，推出适配新模型的Harness 0.2.1代理工具，搜索API今年11月1日前调用免费。

The latest model update brings significant performance improvements and cost advantages, while also opening access to self-developed search and web crawling capabilities for industry users.

The upgraded model, named V4.1-Flash-1017, is updated from the previous V4.1 Flash version. It features a parameter size of 763 billion, with all 11 benchmark tests showing noticeable performance gains compared to the old version. The model is now fully available via web, mobile app and API, with its weights released for public use.

Alongside the model update, two supporting tools have been launched: an API that provides access to self-developed search functions, including Search and Fetch interfaces, priced at 5 yuan per thousand calls; and a new version of the agent tool Harness 0.2.1, specifically optimized for the new model.

When tested on the DeepSWE v1.1 benchmark, the new model demonstrates strong cost efficiency. At its highest performance setting, it achieves a 77.0% pass rate, with each task costing about 1.3 yuan. For lower performance requirements, the low setting offers a 72.6% pass rate at less than 0.8 yuan per task. This is a clear improvement over the previous version, which had a 74.2% pass rate at 3.3 yuan per task, and is far more affordable than competing tools where similar tasks cost nearly 80 yuan each.

The update focuses on optimizing three platforms: its own Harness tool, as well as third-party platforms like CodeBuddy and WorkBuddy. The Harness tool, which added a desktop version last month, needs to be upgraded to version 0.2.1 or higher to fully utilize the new model’s capabilities. New users of Harness get a 6-yuan bonus for their first login, no recharge required, valid for seven days after issuance.

The model also enables "prompt-driven project building", allowing users to create complete projects with fewer prompts and reference materials through interactions in the Harness tool. Trained via self-feedback reinforcement learning, this capability excels at building projects from scratch, including game development and 3D modeling. Examples include a web-based racing game, a Godot-engine remake of *Portal 2*, a valley simulation spanning from prehistoric to modern times, and a 3D model of the Palace Museum.

Another major update is the public launch of its self-developed search engine. This AI-native general search engine, which has long powered the platform’s web search and API functions, handles hundreds of millions of global requests daily, indexes hundreds of billions of web pages, and filters out low-quality content, SEO manipulations and other invalid results. The new Search and Fetch APIs are now available, priced at 5 yuan per thousand calls, with usage instructions in the API documentation. For users integrating the API into tools like Harness, Claude Code or Codex, these interfaces are already in use automatically, with no charges until November 1st.

The model’s weights are now available on public platforms under the name DeepSeek-V4.1-Flash-1017, with MIT open-source licensing. It supports interleaved text and image input, offers a 1 million token context window, and recommended parameters include temperature 1.0, top_p 1.0, and a minimum of 256K max tokens. The API remains accessible under the existing model name "deepseek-flash" with unchanged pricing and call methods.

以上内容不代表本平台立场，仅供读者参考