DeepSeek has officially launched DeepSeek-V4.1-Flash-1017, an updated post-training checkpoint of its V4.1-Flash Mixture-of-Experts architecture. Available through its web portal, mobile app, and API endpoints, the model introduces reinforced learning with self-feedback (RLSF) targeted at end-to-end multi-file software engineering, game creation, and 3D modeling tasks. DeepSeek claims the update delivers enhanced task performance and shorter generation outputs.
According to internal evaluations, the 1017 checkpoint leads benchmarks on DeepSWE, Automation-Bench, and Agents’ Last Exam, while rival frontier models retain leads on Terminal-Bench 4.0 and ProgramBench. API pricing for the model is set at peak rates of $0.30 per 1M cache-miss input tokens and $1.20 per 1M output tokens, with off-peak usage discounted by 50 percent.
Alongside the model checkpoint, DeepSeek opened its Search and Fetch endpoints to public developers. Built on top of DeepSeek’s internal web crawling infrastructure, the endpoints return structured search results or clean Markdown web pages to provide external web context directly to agent applications.
Why it matters
Offers highly competitive API pricing for multimodal MoE models across peak and off-peak execution windows.
Provides integrated web search and page fetch tooling designed to streamline external context retrieval for autonomous agents.
Demonstrates continued progression in self-feedback reinforcement learning methods for code generation and technical agent tasks.
Source: marktechpost.com



