DeepSeek Upgrades V4.1 Flash and Opens Search API

By Suad Seferi · 17 October 2026 at 18:31

DeepSeek has released an upgraded version of its V4.1 Flash artificial intelligence model and opened its proprietary web search infrastructure to developers. The October 17 announcement introduces DeepSeek-V4.1-Flash-1017, with improvements to coding and autonomous task execution, alongside two new APIs for retrieving information from the web. The updated model is already available through DeepSeek's website, mobile application and API. Existing developers can access it using the deepseek-flash model identifier without changing their integration or paying higher model rates. The simultaneous release of web search capabilities gives developers another option for building AI assistants and applications that need current information rather than relying exclusively on knowledge stored during model training. A faster model built for coding and AI agents DeepSeek says V4.1 Flash 1017 completes tasks faster, produces shorter outputs and performs better across several technical evaluations than the previous version released in September. The improvements focus particularly on AI agents, systems that can use tools, execute instructions and complete sequences of tasks with limited human intervention. According to the company's published benchmarks, the updated model achieved a score of 42.3 on Terminal-Bench 4.0, compared with 31.2 for the previous release. Its ProgramBench score increased from 20.3 to 29.1, while its DeepSWE v1.1 result rose from 74.2 to 77.0. These are company-reported results, rather than independent assessments of the new release. Performance in real development environments may differ depending on the tools, evaluation framework and complexity of the task. DeepSeek has also optimized the model for agent frameworks including DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. Developers using DeepSeek Harness are advised to upgrade to version 0.2.1-rc.1 or later. From coding assistance to complete projects One area DeepSeek is emphasizing is the development of complete software projects from relatively short instructions. The company says it has used reinforcement learning with self-feedback to improve the model's ability to handle longer development workflows, including game creation and 3D modelling. In its release materials, DeepSeek demonstrated several projects developed with the model, including a browser-based racing game, a recreation of Portal 2 using the Godot game engine, a 3D model of Beijing's Forbidden City and an interactive historical scene called Valley of Ages. The examples illustrate the type of extended work the model is being trained to handle. They do not, by themselves, establish how consistently it performs across unfamiliar projects or production software environments. DeepSeek opens its own web search engine to developers Alongside the model upgrade, DeepSeek has launched public access to its Search and Fetch APIs. The Search API allows applications to query the web and receive structured results. The Fetch API retrieves the content of an individual webpage and returns it in Markdown format, making it easier for AI systems to process and summarize information. According to DeepSeek, its internal search infrastructure already handles hundreds of millions of requests daily and indexes hundreds of billions of webpages. The company says its search engine supports multiple languages and uses filtering mechanisms intended to reduce search-engine spam, manipulation aimed at AI-generated answers and low-quality AI-produced material. These claims have not yet been independently established through publicly available comparative testing. Both interfaces can be used independently of the language model through authenticated API requests. They are particularly relevant to AI agents, research assistants and retrieval-augmented generation applications that need access to information published after a model's training period. Search pricing starts at $0.75 per 1,000 requests DeepSeek has priced both Search and Fetch at $0.75 per 1,000 successful requests. Charges apply to requests that return usable results or webpage content. Failed requests, timeouts and searches without results are not billed under the published rules. The company is temporarily waiving these search charges for eligible server-side tool calls made through DeepSeek Harness, Claude Code and Codex using the DeepSeek API. Billing for those calls is scheduled to begin on October 31 at 16:00 UTC. The model itself retains its existing API pricing structure, including lower off-peak rates. Developers already using deepseek-flash automatically receive the upgraded model. That distinction matters for development costs. Model inference continues to be billed according to token usage, while the new search tools are charged separately per successful request. Model weights are also available DeepSeek has published the V4.1 Flash 1017 model weights on Hugging Face and ModelScope, allowing developers and researchers to examine the release and explore deployments beyond the company's hosted API, subject to the applicable model terms and infrastructure requirements. The model builds on the V4.1 Flash architecture introduced in September, which combines a mixture-of-experts design with native visual understanding. Publishing the updated weights gives developers a way to evaluate the model separately from DeepSeek's hosted service. However, the Search and Fetch APIs remain distinct hosted services, and releasing model weights does not mean the company's proprietary search index has been made publicly available. For organizations building coding assistants or web-connected AI tools, the October release introduces two practical changes: an upgraded model available through an existing API identifier and a separately priced route to live web information. The next test will be how the upgraded model performs under independent evaluation and whether DeepSeek's search service can consistently deliver the relevance, freshness…