The DeepSeek team officially released the updated V4.1 Flash version, code-named 1017, on October 17, 2026, alongside the Search and Fetch API duo designed to help AI agent systems interact with real-world web data.
The upgrade is now available on the company’s website, mobile app, and API system. At the same time, DeepSeek continues its open-source tradition by releasing the full model weights of the new version on both Hugging Face and ModelScope under a completely free MIT license.
Enhancing goal-oriented programming capabilities
According to DeepSeek’s technical documentation, the V4.1 Flash 1017 release is fine-tuned for wish-driven programming. The model combines reinforcement learning with self-feedback to optimize its ability to build complex end-to-end projects without requiring continuous developer intervention.
The company demonstrated this capability through real-world examples, including building a three-dimensional model of the Forbidden City, creating a racing game that runs directly in the browser, and recreating the game Portal 2 on the Godot Engine.
In terms of benchmark performance, DeepSeek stated that the model achieved superior scores on the DeepSWE v1.1 Pass@1 benchmark when tested via the mini-swe-v2 harness, while also shortening response lengths and reducing task completion times. However, these are self-assessed figures from DeepSeek’s internal environment, and the tech community still awaits independent evaluations to verify its stability when handling real-world source code.
Expanding infrastructure with Search and Fetch APIs
The second key highlight in this announcement is the public rollout of two application programming interfaces: the Search API and Fetch API. Both are powered by DeepSeek’s proprietary in-house search engine, which currently processes hundreds of millions of requests daily worldwide.
The published pricing for each endpoint is set at 0.75 USD per 1,000 calls. To encourage workflow integration among developers, DeepSeek has decided to offer all Search and Fetch calls for free when accessed through server-side tools in DeepSeek Harness, Claude Code, and Codex until 16:00 UTC on October 31, 2026.
The move to enhance both the model and the web browsing tooling layer indicates that DeepSeek is making a decisive shift from a pure language model provider to a comprehensive infrastructure solutions provider for autonomous AI agents.