DeepSeek released DeepSeek-V4.1-Flash-1017 on 17 October 2026, an upgrade to the V4.1 Flash AI model it launched on 10 September 2026. DeepSeek says the new version scores higher on its own coding and agent tests, finishes tasks faster and writes shorter answers. It replaced the previous model on DeepSeek's website, app and API on the same day, at the same price, so any business tool connected to DeepSeek Flash is already running it.
What happened
DeepSeek announced the update in a news post on its API documentation site titled "DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real". The post says the model has stronger benchmark performance, faster task completion and shorter outputs than the version before it, and that it is live on DeepSeek's website, app and API.
The API change log entry of 17 October 2026 adds the detail that matters most to anyone with software already connected. The API (application programming interface, the service other software uses to send requests to the model) is called the same way as before, the pricing is unchanged, and the model name deepseek-flash now points to the new version.
DeepSeek also published the model's weights, the trained files that make up the model itself. The model card on Hugging Face calls the release a "minor update to DeepSeek-V4.1-Flash" that "reduces task execution time and output length", and says the repository and the weights are licensed under the MIT License, a permissive open-source licence.
The same news post opened DeepSeek's Search and Fetch APIs to everyone. They run on DeepSeek's in-house search engine and are sold to developers who want their software to search the web. DeepSeek says calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free for its API users until billing starts on 31 October 2026 at 16:00 UTC.
SiliconANGLE reported on 18 October 2026 that the update arrived a little over five weeks after the original V4.1-Flash launch, and that DeepSeek did not quantify the speed-up.
Key details
| Item | What DeepSeek states |
|---|---|
| Released | 17 October 2026, on the website, app and API |
| API model name | deepseek-flash, now served by DeepSeek-V4.1-Flash-1017; call method unchanged |
| Older names | deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired but still accepted, and are answered by the new model at the Flash price |
| Price per 1M tokens | Off-peak US$0.15 input (cache miss) and US$0.6 output; peak US$0.3 and US$1.2 |
| Peak hours | 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays |
| Context and output | A 1M-token context and output of up to 384K tokens |
| Own test results | DeepSWE v1.1 77.0 (was 74.2), Terminal-Bench 4.0 42.3 (was 31.2), HLE without tools 42.6 (was 36.8) |
| Weights | On Hugging Face under the MIT License |
| Search and Fetch APIs | US$0.75 per 1,000 calls per endpoint; calls through DeepSeek Harness, Claude Code and Codex billed from 31 October 2026, 16:00 UTC |
A token is a small chunk of text; AI services charge per token. The prices come from DeepSeek's pricing page. Every benchmark figure is DeepSeek's own and has not been checked independently.
Why it matters
The biggest change for most businesses is not a test score. It is that a model can change underneath software you already pay for, with no action from you and no new model name to notice. DeepSeek's model card says the update reduces output length. Shorter answers usually mean fewer tokens and a smaller bill per task, but they can also mean a summary that leaves out a line your staff relied on, or an automated reply that reads differently from last week.
Our report on GPT-4o and GPT-4.1 leaving OpenAI's API showed the other version of it: a model removed on a set date. Here, a model is swapped in place. Either way, the AI part of a business tool is a component that moves, and it needs an owner who notices.
The second reason is where the data goes. DeepSeek's privacy policy says that, to provide its services, it directly collects, processes and stores personal data in the People's Republic of China. Under the Australian Privacy Principles, APP 8 covers the steps a business bound by them must take before personal information is disclosed overseas, and its privacy policy has to say whether it is likely to disclose personal information overseas.
Australian Government agencies are already out. The Protective Security Policy Framework lists Direction 001-2025, published 4 February 2025, which requires Australian Government entities to prevent the access, use or installation of DeepSeek products, applications and web services and to remove them from government systems and devices. That direction does not apply to private businesses, but it matters if you supply government clients.
The open weights change the picture. Because the model can be downloaded under the MIT License and run with common serving software such as vLLM and SGLang, a business or its software provider can host it on servers of its own choosing, including an Australian cloud region, so prompts never reach DeepSeek's own service. That takes specialist hardware and skills, and it moves the job of securing and updating the model onto whoever runs it.
DeepSeek also charges double at peak times, set in UTC; our piece on DeepSeek's peak and off-peak pricing explains how that works.
What this means for businesses
Most small and medium businesses do not call DeepSeek's API themselves. The question is whether a tool you use, or one built for you, does. A short checklist:
- Find out who uses it. Ask your software providers and your developer whether any feature sends data to DeepSeek, by name or through a gateway, and whether staff use the DeepSeek app or website for work.
- Re-test what matters. If something relies on deepseek-flash, run a few everyday tasks again: quotes, email replies, document summaries, data checks. Look for answers that are now shorter or missing details.
- Check the data. If customer or staff details are sent, compare that with your privacy policy and APP 8. Where the data is sensitive, ask whether a model hosted in Australia would do the same job.
- Note the Search API date. If a developer has tried DeepSeek's Search or Fetch APIs through a coding tool, those calls are billed from 31 October 2026.
- Name an owner. Someone should keep a short list of the AI models your systems use and read their providers' change logs, so the next silent swap is not a surprise.
If an AI feature sits inside a larger system, such as a CRM linked to your accounting software, treat the model as part of your business systems and integrations, with testing whenever it changes. For a wider review of which AI services your business uses and where its data goes, our IT consulting work starts with that inventory. You can ask us for a free first consultation; we give a written quote before any work starts.
Frequently asked questions
Do I need to change anything if my software already uses DeepSeek Flash?
Usually not in the code. DeepSeek says the API call method and pricing are unchanged and the model name deepseek-flash now points to DeepSeek-V4.1-Flash-1017, so connected software moved over on its own. What is worth doing is re-testing the tasks that matter, because answers may now be shorter and phrased differently.
Is DeepSeek-V4.1-Flash-1017 better than other AI models?
DeepSeek reports higher scores than its September version, for example 77.0 against 74.2 on DeepSWE v1.1, but these are its own results. The comparison that counts for your business is a test on your own documents and tasks, run side by side with the model you use now.
Where does DeepSeek store the data I send it?
DeepSeek's privacy policy says it directly collects, processes and stores personal data in the People's Republic of China. If staff or software send customer or employee details to DeepSeek's app or API, check that against your privacy obligations and your privacy policy before you continue.
Can an Australian business use DeepSeek?
There is no general ban on businesses. Direction 001-2025 applies to Australian Government entities and their systems and devices. If you supply government clients, check your contracts and any security requirements they pass on to suppliers before using DeepSeek services in that work.
What do open weights under the MIT License mean for a small business?
The model files are published for anyone to download and run on hardware or cloud servers they choose, under a permissive licence. In that set-up your data goes to servers you control rather than DeepSeek's, but the model is large and running it needs specialist hardware and skills.
Key takeaways
- DeepSeek released DeepSeek-V4.1-Flash-1017 on 17 October 2026, replacing V4.1 Flash on its website, app and API at the same price.
- Software that calls deepseek-flash switched automatically, so re-test the tasks you depend on, especially where answers are now shorter.
- DeepSeek's own test scores went up, but they are not independent results.
- DeepSeek stores personal data in the People's Republic of China; check that against APP 8 and your privacy policy.
- The MIT-licensed open weights allow self-hosting, which keeps data away from DeepSeek's service but needs specialist skills.