DeepSeek V4.1-Flash-1017: 5 yuan per 1000 requests and "wish-based" coding
DeepSeek V4.1-Flash-1017: 5 yuan per 1000 requests and "wish-based" coding 💡 DeepSeek has officially released the new model DeepSeek-V4.1-Flash-1017, coinciding with the launch of DeepSeek Harness version 0.2.1. The update features specific optimizations for the Harness environment, resulting in faster task execution, reduced output consumption, and lower token costs. To celebrate, eligible users can claim a 6 CNY bonus upon their first login to DeepSeek Harness, valid for 7 days. This limited-time offer is available until the funds run out, providing immediate value for developers testing the new capabilities. ⚡ The new model achieves a "Pareto improvement" in performance and cost. While benchmark scores for coding and Agent tasks have improved significantly compared to the previous V4.1 Flash, the execution time and output length have simultaneously decreased. The model is trained to minimize unnecessary steps and verbose outputs, focusing directly on task completion. In DeepSWE tests, V4.1-Flash-1017 pushes the boundary between performance and cost further out, particularly within DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy frameworks. Users must upgrade DeepSeek Harness to version 0.2.1-rc.1 or higher to fully utilize these enhancements. 🤔 A standout feature is "wish-based programming," powered by self-feedback reinforcement learning. Users only need to describe the desired outcome, and the model autonomously handles planning, coding, execution, and debugging. The model checks its own output and iterates on errors before finalizing the result. DeepSeek demonstrated this by generating a real-time, interactive 3D scene of the Forbidden City, a web-based racing game, and a playable Godot port of Portal 2. This capability transforms simple prompts into complete, functional projects, moving beyond basic code generation to full application development. 📌 What does this mean in practice: Developers can now build complex, interactive applications from single natural language prompts without manually specifying technical stacks or structures. The combination of reduced latency, lower costs, and autonomous debugging makes high-quality AI-assisted development accessible for a wider range of projects, from simple tools to 3D visualizations. 🚀 Additionally, DeepSeek has opened its proprietary search API, featuring Search and Fetch endpoints. The Search API allows agents to find web resources with filters for site, language, and date, while the Fetch API converts selected pages into Markdown for analysis. The pricing is extremely competitive at 5 CNY per 1,000 successful requests, with no charge for failed or timed-out queries. This API is currently free until November 1st and can be integrated into DeepSeek Harness, Claude Code, or Codex, further solidifying DeepSeek's position as a cost-effective leader in the AI ecosystem. Перевод и адаптация: AI Lab. Источник: QbitAI (量子位), https://www.qbitai.com/2026/10/504067.html #AI #DeepSeek #LLM #Coding #API #TechNews