Written by: Fan Jialai
Source: The Paper
"A big one is coming." A few days ago, DeepSeek's sudden outage led many AI enthusiasts to speculate that the leading domestic large model company DeepSeek was about to release another major update—and as expected, it really did arrive.
On October 17, The Paper reporter learned that the leading domestic large model company DeepSeek has released another version update: DeepSeekV4.1 Flash has now been updated to version V4.1-Flash-1017, and the official web, APP, and API models have all been switched to the new model.
According to the official introduction, compared with the previous version of DeepSeekV4.1 Flash, the new model has achieved substantial improvements in benchmark tests across various aspects.
While improving performance, it shortens the model's task runtime and output length, enabling users to complete tasks at lower cost and faster speed, achieving stronger cost-effectiveness under multiple Harnesses. (Officially called a "performance-cost Pareto improvement.")
The so-called Harness refers to a software support framework built around a large language model, designed to help AI Agents manage and execute more complex tasks. Simply put, it can be understood as the entire runtime system needed for a large model to truly become an Agent.
It is worth noting that the DeepSeek Harness team also posted on the same day: the DeepSeek Harness 0.2.1 version released today is adapted to the newly released DeepSeek-V4.1-Flash-1017 model. During training, the new model was specifically optimized for DeepSeek Harness, improving task performance while shortening task runtime and reducing output consumption and token costs.
First-time login and use of DeepSeek Harness will grant 6 yuan in bonus credit, with no recharge required. The bonus credit is issued in limited quantities, available while supplies last, and valid for 7 days after arrival.
In terms of capabilities, DeepSeekV4.1 Flash 1017, through reinforcement learning (RL) training technology based on self-feedback, enhances 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. It also focuses on optimization for platform frameworks such as DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.
"Wish-style" programming means users can describe development requirements in natural language, and AI autonomously completes the development, reducing manual intervention.
Forbidden City 3D modeling display effect
Web-based racing game development effect
DeepSeek also announced the opening of its self-developed search API. According to the introduction, DeepSeek Search is an AI-native general-purpose search engine that indexes hundreds of billions of web pages worldwide, supports multilingual search, and filters SEO cheating, GEO manipulation, and low-quality AIGC through strict identification mechanisms, ensuring that authoritative and trustworthy information is presented to users and LLMs.
In terms of pricing, this update can be called the king of cost-effectiveness: according to the introduction, this new version release of the DeepSeek API adds support for the Search and Fetch interfaces, and the related capabilities are officially open to the public. The price for both interfaces is 5 yuan per thousand calls.
At present, when users use the DeepSeek API in DeepSeek Harness, Claude Code, or Codex, they will already access the Search/Fetch interfaces through server-side tool calls, and it is free for a limited time until November 1.