On October 17, 2026, DeepSeek released the V4.1-Flash-1017 large language model. According to an announcement on the company's official website, the V4.1 Flash series has been updated to checkpoint 1017: project-building capabilities are significantly stronger, usage costs are lower, and tasks run faster. Alongside the update, Search and Fetch API support was added for web search and page collection; the company invited developers to try the new capabilities and share feedback. The announcement was published on October 17 on the company's homepage. Earlier news about the company is available in the Aishunos archive.
Key highlights of the update
DeepSeek V4.1-Flash-1017 is a minor update to V4.1 Flash that, according to the model card, delivers notable gains in reasoning, coding, agentic tasks, and visual benchmarks. The model weights were published on Hugging Face under the MIT license. The license opens the model to researchers, startups, and enterprises: they can download it, run it on their own infrastructure, modify it, and ship it in commercial products.
According to the official announcement, checkpoint 1017 significantly strengthens V4.1 Flash's project-building capabilities, lowers usage costs, and speeds up task execution. The company highlights these three metrics as the update's main achievement.
The official announcement emphasizes the update's three pillars: significantly stronger project-building capabilities, lower usage costs, and faster task execution. The company describes this as higher performance without a higher price — being stronger and cheaper at the same time.
The company calls the update a Pareto improvement: performance increased while the price stayed the same. The official statement says benchmark scores improved, tasks complete faster, and outputs are shorter. API pricing is unchanged, and developers access the model through the existing deepseek-flash endpoint.
"This update is a Pareto improvement in performance and price: stronger benchmark results, faster task completion, and shorter outputs; API pricing is unchanged, and developers access the model through the existing deepseek-flash endpoint," the company said.
Benchmark results
According to results published in the model card, V4.1-Flash-1017 scored 77.0 on DeepSWE v1.1 for software engineering — above the previous Flash version's 74.2. Terminal-Bench 4.0 recorded 42.3 points, and the tool-free HLE test recorded 42.6 points.
These numbers show the update is aimed squarely at practical tasks: DeepSWE measures software engineering, Terminal-Bench measures terminal-environment performance, and HLE measures general knowledge and reasoning. The DeepSWE gain (from 74.2 to 77.0) versus the previous Flash version is considered notable for a minor update.
These are the official results published in the model card and should not be confused with independent evaluation results. Even so, the observed improvement over the previous Flash version demonstrates the update's practical effectiveness.
Search and Fetch API
Alongside the update, DeepSeek introduced two new interfaces: Search and Fetch. They are powered by the company's own search engine and designed for AI agents and RAG applications. Through Search the model accesses web search results, while Fetch enables page content collection (scraping) — an agent can open a web page and read its text directly. The new interfaces are covered in detail in the official API documentation.
According to the official documentation, these interfaces are designed specifically for agents and RAG applications: the model can fetch up-to-date web data directly before forming a response. That is a meaningful convenience for applications working with constantly changing data.
The service is billed per request: $0.75 per 1,000 successful requests — only successfully completed requests are charged. Through DeepSeek Harness, Claude Code, or Codex integrations, usage is free until November 1, 2026 — opening the door for developers to try the new capabilities at no cost.
Pricing and usage
API pricing on DeepSeek's own servers is unchanged: $0.15 per 1 million input tokens and $0.60 per 1 million output tokens. The context window is 1,048,576 tokens.
Unchanged pricing is one of the update's headline messages: the company raised performance while holding the price. Nothing changes for existing integrations — requests go to the deepseek-flash endpoint as before, with no new address to configure. And the 1,048,576-token context window makes it possible to process large codebases and long documents in a single request.
