In this issue
The short version
- DeepSeek open-sourced DeepInfer, the inference framework behind its own service, under Apache 2.0. It supports Nvidia Blackwell and Huawei Ascend chips and currently runs V4.1-Flash-1017.
- Ant Group's Ming-3.0 speech series has five open models for speech synthesis, audio understanding and speech recognition, already used in Ant's own apps.
- A US DHS order requires civilian federal agencies to stop buying and remove DeepSeek products, including third-party software built on its models. It does not apply to businesses or the public.
- If you work with US government clients, check whether any tool in your stack uses DeepSeek models.
October 21 combined open-source releases from two Chinese companies with a US government order that limits where one of them can be used. Together they show both the appeal of open models and the compliance questions that come with them. SpaceX's third-quarter results, also reported this day, are covered in our October 27 issue.
01
DeepSeek DeepInfer: the engine behind DeepSeek's service
What DeepSeek released
DeepSeek (深度求索) open-sourced DeepInfer on October 20 under the Apache 2.0 licence, describing it as the high-performance inference framework that powers its large-scale production service. It includes distributed inference, speculative decoding and structured output, and supports Nvidia's Blackwell GPUs and Huawei's Ascend 950 (华为昇腾 950) processors; older Nvidia generations such as Hopper are not yet supported, and some features work only on Nvidia hardware. It currently runs DeepSeek-V4.1-Flash-1017, and DeepSeek says future models will follow. A companion component, deepseek-recipe, handles request conversion and input and output processing the way DeepSeek's own service does.
The public release supports offline inference; building an online service on it still needs more work. DeepSeek says the aim is to help developers and companies cut computing costs and energy use.
When self-hosting an open model is worth it
Most teams should keep using a hosted API: it is cheaper than running GPUs that sit idle. Self-hosting makes sense when data cannot leave your infrastructure, when volume is high and steady enough to keep hardware busy, or when you need behaviour that matches the vendor's service exactly. DeepInfer helps with the last point, because it is the same engine DeepSeek runs.
Before deciding, estimate your monthly tokens, compare the API bill with the cost of the required GPUs at realistic utilisation, and add the engineering time to run them. Run a pilot on rented hardware for a week before buying anything. DeepSeek's hosted service is the baseline to beat; our DeepSeek review covers its pricing.
Get HAI Daily in your inbox
The day's AI news, what it changes, and which tools to try.
02
Ant's Ming-3.0 open speech models
What Ant Group released
The Bailing (百灵) team at Ant Group (蚂蚁集团) open-sourced the Ming-3.0 speech series on October 20: five models covering speech synthesis, audio understanding and dialogue, and speech recognition. The speech synthesis models let you describe a voice in text and control emotion, rhythm and emphasis, and can insert sounds such as laughter or sighs. The audio models can transcribe, summarise and analyse long recordings directly, keeping information about emotion, speakers and setting. The small recognition model accepts lists of key terms and conversation context to help with jargon, dialects and mixed Chinese-English speech; Ant reports an average error rate of 4.41% across such tests.
The weights are on Hugging Face and ModelScope (魔搭社区), and Ant says the models already run in its own products, including its health assistant and Alipay (支付宝) search. Performance figures come from Ant.
Choosing a speech model for your project
For most people, hosted voice tools such as ElevenLabs or Fish Audio remain the quickest route to natural speech. Open models like Ming-3.0 matter when you need to process sensitive recordings on your own servers, handle Chinese dialects and mixed-language speech, or avoid per-minute fees at high volume.
Test with your own audio. For recognition, take ten recordings with your real vocabulary, add a list of key terms and count the errors that change the meaning. For synthesis, generate the same script in two tools and ask listeners which sounds more natural. Check the licence on the model card before commercial use, and get consent before cloning anyone's voice. Our ElevenLabs vs alternatives guide compares hosted options, and the STT & TTS category lists more.
03
US DHS orders DeepSeek out of federal systems
What the order requires
The US Department of Homeland Security issued an order under the Federal Acquisition Supply Chain Security Act requiring federal civilian agencies to stop buying DeepSeek products and remove them from their systems. Signed in July and published in October, it covers DeepSeek's own products, its affiliates and third-party software built on its models, whether or not the vendor is American. DHS cited three concerns: data stored on servers in China could be exposed, the model could help gather intelligence against the US, and vulnerabilities in DeepSeek's iOS app.
The order does not apply to businesses or the public, and reports say it excludes Defence Department agencies. Some security experts questioned how much evidence DHS had published to support it. It remains unclear whether gateway services that route requests to DeepSeek models are covered.
What it means for vendors and their customers
If you sell software to US federal civilian agencies, find out now whether any part of your product calls a DeepSeek model or embeds its weights, including through third-party libraries or AI gateways. Document the answer, because customers will ask. Keep a model inventory for every AI feature: which model, which provider and where requests are processed.
For everyone else, the order is a reminder that the country behind a model can affect where you are allowed to use it. If your clients include governments or regulated industries, have a tested alternative such as Qwen (通义千问) or a US-hosted model ready, and check each client's rules before choosing. The AI Security category lists tools that help audit software dependencies.
What to try next
Put one change into practice
Compare your monthly API bill with the true cost of self-hosting before choosing an open model. Test speech models on ten of your own recordings. Build a model inventory listing every AI model your products use and where it runs.
Explore Code Generation, STT & TTS and AI Security to build your shortlist.
How we put this issue together
The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.