DeepSeek, the Hangzhou lab whose cheap, open models keep resetting AI prices, had a busy week. On 17 October it shipped an update to its main model, V4.1-Flash-1017. On 20 October it open-sourced DeepInfer, the in-house software that serves its models to millions of users. And on the same day, the US Federal Register published an order telling federal civilian agencies to remove DeepSeek technology from their systems, including inside third-party products.
DeepSeek's week
Order signed
US Homeland Security signs the exclusion order, unpublished at the time.
V4.1-Flash-1017
Faster model update, same API name and price; Search and Fetch APIs launch.
Databricks switches
Databricks adds the 1017 version and schedules the original V4.1 Flash for retirement on 19 November.
DeepInfer open-sourced
The serving engine behind V4.1 Flash, on GitHub under Apache 2.0.
Order publishedNow
The Federal Register publishes the DHS order; China's foreign ministry objects.
V4.1-Flash-1017: a quiet but useful update
The 1017 release keeps the same API name and price as the September V4.1 Flash, so existing users get it automatically. It answers faster and more briefly, and scores better on agent-style coding. The weights are on Hugging Face and ModelScope under the MIT licence, with the same 1-million-token context.
V4.1 FlashV4.1-Flash-1017
Higher is better. DeepSeek-reported scores.
DeepSeek also made two web tools generally available, Search and Fetch, so its models can look things up during a task. They are priced per call rather than per token, and DeepSeek is offering free server-side use inside some coding tools until the end of October.
DeepInfer: the serving engine, given away
Training a model gets the headlines, but serving it cheaply is where AI companies make or lose money, and they usually keep those techniques secret. DeepInfer is the framework DeepSeek uses to run V4.1 Flash in production. Releasing it means cloud providers and companies that self-host can now use the same tricks.
- What it is
- DeepSeek's production inference framework, on GitHub as deepseek-ai/DeepInfer
- Licence
- Apache 2.0
- Models supported
- DeepSeek-V4.1-Flash only, for now
- Hardware
- Nvidia SM100 data-centre GPUs and Huawei Ascend 950 chips
- Included
- optimised model code, multi-chip parallelism, speculative decoding, consistent outputs between runs, structured output (regex, JSON, grammars)
- Not included
- caching system, separate prompt and generation servers, load balancing across machines
- Built on
- mainly vLLM, with ideas from SGLang and Nvidia TensorRT-LLM
It caught on fast: the repository passed 8,000 GitHub stars within a day, and the vLLM project said it will borrow ideas from DeepInfer to improve its own support for V4.1 Flash. The Huawei support matters too. DeepSeek has said it plans to train on Huawei chips by the end of 2026, and an open, production-grade engine for Ascend hardware makes Chinese chips a more realistic choice for running models.
The US order: what it covers
The order uses the Federal Acquisition Supply Chain Security Act of 2018, a law first used in 2025. It was signed by Homeland Security Secretary Markwayne Mullin on 24 July, filed for public inspection on 19 October and published on 20 October. The published notice gives no reasons; DeepSeek's privacy policy says it stores users' personal data in China, which has been the main concern behind earlier device bans.
Covered
federal civilian agencies- Exceptions
- possible, by written request with a compelling justification
- Remove DeepSeek products from their IT systems
- Remove third-party products or services "developed using DeepSeek implementations"
- No new purchases from DeepSeek or its affiliates
Federal contractors
through FAR clause 52.204-30- May not provide or use covered products in a federal contract without a waiver
- Must check SAM.gov for new orders at least every three months
- Must report covered products within 3 business days, with mitigation within 10
Not covered
companies and the public- Private use of DeepSeek is not restricted by this order
- Defence and intelligence agencies fall under separate orders
The phrase "DeepSeek implementations" is not defined, which is what worries software vendors: a product that quietly routes some requests to a DeepSeek model, or uses a model fine-tuned from DeepSeek weights, may be caught. China's foreign ministry said Beijing opposes "overstretching national security" and will protect Chinese companies' rights.
What it means for your team
- 1
List the models in your stack
Include models used by vendors and AI features inside SaaS tools, not just the ones you call directly.
- 2
Check your customers
If you sell to US federal agencies, assume they will ask whether DeepSeek is anywhere in your product.
- 3
Keep a fallback
Design AI features so the model can be swapped, through a gateway or a config setting, without rewriting code.
- 4
Weigh self-hosting
If DeepSeek's price and quality suit you and your rules allow it, open weights plus DeepInfer make running it in your own cloud more practical.
Check your understanding
1/3Under which licence did DeepSeek release DeepInfer?
The model weights are MIT; the DeepInfer engine is Apache 2.0.
2/3Who must follow the US order?
It does not restrict companies or the public.
3/3What changed in V4.1-Flash-1017 for existing API users?
The API name and prices stayed the same.
You got 0 out of 3.
More on open models this month: Nvidia Nemotron 3.5 and Reflection Beam and the best AI models in October 2026. For the wider regulatory picture, read AI rules this month.
Sources: The Next Web and AI Weekly on DeepInfer (20 October 2026); the deepseek-ai/DeepInfer GitHub repository; DeepSeek V4.1-Flash-1017 release coverage (17 October 2026); Databricks release notes (19 October 2026); The Hacker News on the DHS FASCSA order and the Federal Register notice (20 October 2026). Benchmarks are DeepSeek's own. This is not legal advice.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 21 October 2026. About me