DeepSeek Releases Updated V4.1 Flash Model With Cheaper Coding Agents

Chinese firm DeepSeek launched an upgraded Flash model featuring enhanced programming capabilities, faster processing, and lower operating costs.

NationPulse Desk

Oct 17, 2026 · 9:08 PM 3 min read

Key takeaways

DeepSeek released DeepSeek-V4.1-Flash-1017 with upgrades in coding, agentic tasks, speed, and output efficiency.

The new model uses a training method called reinforcement learning with self-feedback, or RLSF.

On the DeepSWE benchmark, V4.1-Flash-1017 scored 77.0 percent at an average cost of under $0.20 per task.

DeepSeek opened its internal web search technology to outside developers through Search and Fetch APIs costing $0.75 per 1,000 calls.

The release arrives as the Chinese government reportedly considers tighter controls on open-weight AI models.

Chinese artificial intelligence developer DeepSeek launched DeepSeek-V4.1-Flash-1017 on its platform, bringing enhancements to programming capabilities, agentic tasks, processing speed, and response efficiency. Users can access the system via the official website, mobile application, and application programming interface, while the firm also published model weights on Hugging Face and ModelScope. According to the organization, the new version achieves higher benchmark results than its earlier iteration while finishing assignments quicker and producing concise answers, which the business calls a Pareto improvement balancing expenses and capabilities.

New Training Method

For complete software engineering work, DeepSeek applies a post-training procedure named reinforcement learning with self-feedback, or RLSF. This strategy lets the software rely on internal confidence ratings or self-generated evaluations as reward signals instead of depending on human reviewers or outside reward frameworks. The developer states this boosts the system proficiency in constructing full projects from the ground up, spanning activities like game creation, three-dimensional modeling, and aesthetic design. The software is also capable of producing compact animated explainer videos. Furthermore, DeepSeek optimized V4.1-Flash-1017 for coding platforms including DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy, suggesting version v0.2.1-rc.1 or newer for DeepSeek Harness.

Smaller Model Beats Larger Rivals

Based on internal testing by DeepSeek, V4.1-Flash-1017 surpasses the older V4.1-Flash version across all evaluated metrics. The organization asserts that the compact model defeats larger competitors such as Claude Opus 5, GPT-5.6 Sol, Kimi K3, and GLM 5.3 on multiple programming and artificial intelligence agent evaluations. On the DeepSWE assessment, V4.1-Flash-1017 attained a maximum score of 77.0 percent with an average expense of under $0.20 per assignment. Operating at a reduced setting, the system still achieved over 72 percent while running near $0.10 per task. The older version achieved 74.2 percent at a price of roughly $0.50 per execution. Charts from the firm position systems like GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 between roughly $2 and exceeding $10 per task while displaying reduced accuracy compared to V4.1-Flash-1017. External evaluations remain absent, and practical developer evaluations concerning code quality might vary from test scores.

Search Engine Opened to Developers

Simultaneously with the model rollout, DeepSeek is making its proprietary web search technology available to external programmers for the first time. The new Search and Fetch application programming interfaces use the company search engine, which the business reports currently processes hundreds of millions of requests daily across the globe. Each endpoint carries a price of $0.75 per 1,000 requests. Users of the DeepSeek application programming interface can use the tools without charge via server-side connections in DeepSeek Harness, Claude Code, and Codex until October 31, 2026, at 4:00 PM UTC.

China Could Restrict Open-Weight Models

This deployment occurs amid reports that authorities in China are weighing stricter oversight regarding the distribution of open-weight artificial intelligence systems and other vital technologies. Under the discussed framework, whether a system ranks among the 10 most powerful technologies could serve as a criterion for deciding if extra regulations take effect. Such measures might complicate heavy dependence on Chinese open-weight systems if stricter publication or distribution mandates are enacted.

Questions and answers

How much does the DeepSeek Search and Fetch API cost?

Each endpoint costs $0.75 per 1,000 calls, though API users can access them for free through specific server-side integrations until October 31, 2026.