Chinese artificial intelligence company DeepSeek released an updated version of its Flash model on October 17, 2026, with improvements aimed at faster task execution, higher efficiency, and shorter outputs. The new model, DeepSeek-V4.1-Flash-1017, is available through the company's website, mobile app, and API, and its weights have been published as open source on Hugging Face and ModelScope. The company described the release as a Pareto improvement in the tradeoff between performance and cost.
Efficiency and Benchmark Gains
DeepSeek reported that the updated model improves on its direct predecessor in every tested category. The company highlighted gains in agentic and coding tasks, with internal benchmarks showing stronger accuracy, faster task completion, and shorter responses. These changes are intended to support applications that require multi-step reasoning and rapid processing of large data volumes.
Lower Costs per Task
In the DeepSWE benchmark, V4.1-Flash-1017 achieved a top score of 77.0% at an average cost of slightly under US$0.20 per task, while a lower setting still cleared 72% for about US$0.10. The predecessor scored 74.2% but cost roughly US$0.50 per run. DeepSeek's cost chart also places rival models such as GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 between US$2 and over US$10 per task without matching the smaller model's accuracy.
Training with Self-Feedback
DeepSeek used reinforcement learning with self-feedback, or RLSF, to improve end-to-end project development. This post-training method relies on the model's internal confidence scores and self-generated critiques as a reward signal, removing the need for human evaluators or external reward models. The company said the approach supports tasks ranging from game programming to 3D modeling and design work, with examples such as a web-based racing game and a recreation of Portal 2.
Search and Fetch APIs
Alongside the model update, DeepSeek opened its in-house search engine to outside developers for the first time. The new Search & Fetch APIs run on infrastructure that already handles hundreds of millions of queries daily worldwide, according to the company. Each endpoint costs US$0.75 per 1,000 calls, while DeepSeek API users can access the features for free through server-side tools in DeepSeek Harness, Claude Code, and Codex until October 31, 2026 at 4:00 PM UTC.
Developer Tools and Competitive Positioning
The model has been tuned for development environments such as DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy, with DeepSeek Harness version v0.2.1-rc.1 or later recommended for the best experience. DeepSeek's benchmarks show the model outperforming larger competing systems, including Claude Opus 5, GPT 5.6-Sol, Kimi K3, and GLM 5.3, in several coding and AI agent categories. The company competes with global and Chinese AI developers on reasoning, programming, tool use, and cost efficiency.
Open Weights and Regulatory Context
DeepSeek published the model weights as open source on Hugging Face and ModelScope, continuing its practice of making models available to researchers and developers. The release arrives as media reports suggest Chinese regulators could tighten rules on publishing open-weight models or even block their release, depending on whether a system ranks among the top 10 most capable models. This backdrop adds strategic weight to DeepSeek's decision to ship the updated model publicly.
The new release highlights DeepSeek's focus on practical efficiency rather than raw model scale alone. Shorter outputs and faster response times could reduce processing overhead for applications that handle large volumes of data or repeated multi-step tasks. Independent benchmarks and developer feedback will ultimately determine how the update performs in real-world use, but the combination of open weights, lower task costs, and new search tools strengthens DeepSeek's position in the commercial AI market.