DeepSeek released another model update on Oct. 17, with DeepSeekV4.1 Flash upgraded to V4.1-Flash-1017 across its web platform, app, and API, according to a report by The Paper.
The company said the new model delivers broad gains in benchmark testing compared with the previous DeepSeekV4.1 Flash release. At the same time, it cuts task runtime and output length, which lowers user costs and speeds up completion. DeepSeek described the result as a "Pareto improvement" between performance and cost across multiple Harness setups.
The report described Harness as a software support framework built around large language models to help AI agents manage and execute more complex tasks. In simpler terms, it is the operating system layer needed to turn a large model into a functioning agent.
DeepSeek Harness said in a same-day post that its newly released version 0.2.1 is compatible with DeepSeek-V4.1-Flash-1017. The team said the new model received targeted optimization for DeepSeek Harness during training, improving task performance while reducing runtime, output consumption, and token costs.
Users logging into DeepSeek Harness for the first time can receive a 6-yuan credit without making a deposit. The credit is available in limited quantities and remains valid for seven days after it is issued.
On capabilities, DeepSeek said V4.1 Flash 1017 uses reinforcement learning based on self-feedback to improve its ability to build complete projects from scratch in areas such as game development and 3D modeling. The report said the model can handle more "wish-based" programming tasks and project development requests, and that it has been optimized for frameworks including DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.
In the report, "wish-based" programming refers to a workflow in which users describe development needs in natural language and the AI completes the development process with less manual intervention.
DeepSeek also said it is opening its self-developed search API. According to the company, DeepSeek Search is an AI-native general search engine that indexes hundreds of billions of webpages worldwide, supports multilingual search, and uses strict filtering mechanisms to screen out SEO abuse, GEO manipulation, and low-quality AIGC content so that users and LLMs receive authoritative and reliable information.
On pricing, the latest DeepSeek API update adds support for Search and Fetch interfaces, making those capabilities publicly available. Both interfaces are priced at 5 yuan per 1,000 calls.
Users can already access the Search and Fetch interfaces through server-side tool calls when using DeepSeek API in DeepSeek Harness, Claude Code, or Codex. That access will be free for a limited time through Nov. 1.