On October 17, 2026, the Chinese AI company DeepSeek released DeepSeek-V4.1-Flash-1017, an updated version of its lightweight model V4.1 Flash. The "1017" in the name refers to the release date. From the same day, DeepSeek's website, app and API switched over to the new version. The company also opened the Search & Fetch API, which lets outside developers use the search engine DeepSeek built in-house, to the public for the first time.
What changed in V4.1-Flash-1017
DeepSeek describes the update as one that "scores higher on evaluations, finishes tasks faster and produces shorter output." It says it carried out additional training using reinforcement learning, a method in which a model improves through trial and error, using its own results as clues.
In the API, the model name stays "deepseek-flash" as before, and it now points to the new version. DeepSeek says the way the API is called and its pricing are unchanged. For the best performance, it recommends using version 0.2.1-rc.1 or later of DeepSeek Harness, its agent software.
The model weights (the data that makes up the trained model) were also released on Hugging Face and ModelScope, under the MIT License, which allows free use, including commercial use. According to the Hugging Face model card, the model has 763 billion (763B) parameters in total, and the amount of text it can take in at once (its context) is 1 million tokens.
Evaluation results
The main results published by DeepSeek are as follows. All were measured by DeepSeek itself, and the models it compared against were OpenAI's GPT-5.6 Sol, Anthropic's Claude Opus 5, Moonshot's Kimi K3 and Z.ai's GLM-5.3.
- DeepSWE v1.1 (software development): 77.0, up from 74.2 for the previous version and the highest among the models compared (Claude Opus 5 scored 74.0 and GPT-5.6 Sol 73.0)
- Terminal-Bench 4.0 (command-line operations): 42.3, up from 31.2 for the previous version, but short of Claude Opus 5's 51.8
- HLE (expert-level hard questions, without tools): 42.6, up from 36.8 for the previous version. Claude Opus 5 scored 56.3 and GPT-5.6 Sol 44.5
- Apex Bench v1.1: 72.1, up from 64.6 for the previous version and the highest among the models compared
DeepSeek says Terminal-Bench 4.0 was measured using the Claude Code framework and DeepSWE v1.1 using a framework called Mini-SWE-V2. While the model came out on top among those compared on some programming evaluations such as DeepSWE, Claude Opus 5 stays ahead on other programming evaluations in the same table, such as ProgramBench (29.1 versus 37.0 for Claude Opus 5) and NL2Repo-Bench (68.2 versus 75.3), as well as on the hard questions and command-line operations.
Search & Fetch API, built on DeepSeek's own search
The Search & Fetch API, opened at the same time, lets developers call the search engine DeepSeek developed in-house from their own apps and AI agents. It has two parts: "Search" for web searches and "Fetch" for retrieving the content of a specified page.
- Pricing: US$0.75 per 1,000 successful requests
- Limits: up to 5 requests per second for both Search and Fetch (Fetch also has per-site limits)
- Free period: free when used as a DeepSeek server-side tool from DeepSeek Harness, Claude Code or Codex, until billing starts at 0:00 on November 1, 2026, Beijing time (1:00 the same day, Japan time). Calling the API directly is charged from the start
DeepSeek has given no specific information about use from Japan or the quality of Japanese-language searches.
CoAI's take
DeepSeek has once again stuck to its approach of improving performance without changing prices and releasing the weights under the MIT License. It says the model beats other companies' top models on some programming evaluations such as DeepSWE, which makes it a strong option for developers who want to hand programming work to AI while keeping costs down. Opening its own search API can be seen as a move to let developers build AI agents with built-in search entirely on one company's services.
On the other hand, the evaluation figures were measured by DeepSeek itself. There are gaps on other programming evaluations, hard questions and command-line operations, so for business use it is best to decide after testing it on your own work. If you use DeepSeek's API, you should also check its terms of use, including where the data you send is handled.
Sources
- DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real (DeepSeek API Docs)
- Change Log (DeepSeek API Docs)
- Search & Fetch (DeepSeek API Docs)
- deepseek-ai/DeepSeek-V4.1-Flash-1017 (Hugging Face)
- Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs (The Decoder)
