QWE AI Academy Here’s a detail buried in the Reuters scoop that nobody’s talking about: DeepSeek, CXMT and the other firms were already approved for the Entity List last year. The blacklisting is sitting in a drawer. Reuters’ June 16, 2026 report confirmed the US has held off adding DeepSeek and 100+ other companies to the trade blacklist while the Trump administration tries to avoid escalating tensions with Beijing. The decision was made. The publication wasn’t.
That’s a weird position for any team relying on DeepSeek’s API. The risk hasn’t gone away – it’s been deferred. If you’re using DeepSeek for anything that matters, this post is a practical playbook: what the news actually changes, what it doesn’t, and how to set up DeepSeek so the next headline can’t break your workflow.
The problem: the headline buys you maybe six months
Most coverage is framing this as good news for DeepSeek users. It isn’t. The interagency committee already signed off – the only thing keeping DeepSeek off the Entity List is trade-deal optics. Per the same Reuters reporting, a senior State Department official told Reuters that DeepSeek has supported China’s military and intelligence operations and tried to use Southeast Asian shell companies to obtain US chips. That assessment doesn’t get walked back. It just gets timed.
If you wake up one morning to find DeepSeek listed, the practical consequence is brutal: US companies can’t ship goods, software or technology to listed firms without a license, and those licenses are typically denied. Any cloud integration, payment processor, or US-based dev tool sitting between you and the DeepSeek API would have to cut you off. Overnight.
Why the obvious solutions don’t work
Two reactions show up everywhere on Reddit and X right now. Both are wrong for most teams.
“Just use the DeepSeek API while it’s still up.” The API sends your prompts to servers in China. Even setting aside geopolitics, Anthropic published a detailed report on February 23, 2026 accusing DeepSeek (along with Moonshot and MiniMax) of running industrial-scale distillation campaigns against Claude – 16 million exchanges through 24,000 fake accounts. Whatever your view of that fight, the company you’re shipping prompts to is mid-lawsuit-grade dispute with two of the biggest US AI labs.
“Switch to Claude or GPT-4.” Fine for chat. Painful if you picked DeepSeek for the reasoning quality at a fraction of the price, or for the open weights. You’d be giving up the actual reason you chose it.
There’s a third option almost no tutorial in the news cycle is pointing at: the DeepSeek API and the DeepSeek model are two different products. The weights are MIT-licensed and sit on your hardware. The export-control mess applies to the company, not to a file on your disk.
The recommended approach: run the weights locally
Stop using the hosted API. Run an open-weight DeepSeek variant through Ollama instead – your prompts never touch a server you don’t control. There’s no ongoing transaction with DeepSeek the company, just a model file you already downloaded. The Entity List question stops being your problem the moment you make that switch.
One thing every “run DeepSeek locally” guide glosses over: the full R1 is a 671 billion parameter Mixture-of-Experts model released January 20, 2025, and running it requires roughly 1.5 TB of VRAM. Nobody’s running that on a MacBook. What you’re actually running locally is a distilled version – a smaller Qwen or Llama base trained to mimic R1’s reasoning. Useful. Not the same model. Be honest with yourself about that.
The three-command setup
On Linux or macOS, the install is genuinely this short:
curl -fsSL https://ollama.com/install.sh | sh
ollama pull deepseek-r1:7b
ollama run deepseek-r1:7bThe 7B distilled model runs on an M-series Mac with 16GB+ unified memory or any GPU with ~8GB VRAM (discrete, not shared). For the 32B variant on cloud, a g5.2xlarge on AWS with its 24GB GPU is the usual pick – running at $1.212/hour on-demand as of mid-2026, though you should check current AWS pricing before committing. Smaller models down to 1.5B will run on a regular laptop CPU, slowly.
There’s something worth sitting with here: “offline AI” sounds like a marketing promise, but with Ollama it’s a verifiable fact. You can disconnect your ethernet cable, run a query, and watch it complete. That’s a different category of assurance from a vendor telling you data stays private on their servers.
Verify it’s actually offline
This is the step every guide skips. Open a network monitor (Little Snitch on macOS, nethogs on Linux), then run a query. Zero outbound traffic from the Ollama process means the model isn’t phoning anywhere. The Chinese-data-law concern becomes structurally impossible – not promised away, but mechanically ruled out. Belt-and-suspenders option: run the whole thing inside a container with no outbound network access.
What this looks like in a real workflow
A 4-person legal-tech team was using the DeepSeek API for clause extraction on contract PDFs – about 2 million tokens a day. The bigger win for them: the partner contract review that had been blocking them for weeks – a client requirement that no data leave a US-controlled environment – went through immediately once they moved off the hosted API. They shifted the pipeline to a single g5.2xlarge running deepseek-r1:32b via Ollama’s OpenAI-compatible endpoint. Their existing code needed exactly one change: the base URL.
from openai import OpenAI
client = OpenAI(
base_url="http://your-ec2-ip:11434/v1",
api_key="ollama" # ignored, but the SDK requires a string
)
resp = client.chat.completions.create(
model="deepseek-r1:32b",
messages=[{"role": "user", "content": "Extract all indemnity clauses..."}]
)Monthly spend went from variable API costs to a flat ~$870/month for the EC2 instance running 24/7 (at mid-2026 on-demand pricing – schedule it and that drops substantially).
Pro tips from people who’ve already hit walls
Pro tip: Don’t treat the distilled 7B/8B models as a drop-in for the API’s reasoning quality. The hosted DeepSeek-R1 is the full 671B MoE. The 7B distill is built on Qwen, the 70B distill is built on Llama. They inherit some of R1’s chain-of-thought style but lose real accuracy on hard math and multi-step coding. For anything requiring deep reasoning, jump to 32B or rent the 70B on cloud – don’t try to make 7B do work it can’t do.
- Pin your model version. Ollama updates tags silently. If
deepseek-r1:32bgets re-pulled with new weights, your eval numbers can shift overnight. Use the specific quantization tag (e.g.deepseek-r1:32b-qwen-distill-q4_K_M) and document it. - Watch the context window. Distilled DeepSeek variants on Ollama default to a much smaller context than the API. Set
num_ctxin the Modelfile if you’re feeding it long documents – otherwise it’ll silently truncate. - The legitimacy question is unresolved. Anthropic’s own report admits distillation is a widely used and legitimate training method – labs routinely distill their own models. The accusation against DeepSeek isn’t that distillation happened; it’s that it happened through fake accounts against ToS. Worth knowing if a stakeholder asks whether you’re using “stolen” weights. You aren’t. You’re using weights released under MIT.
What to do this week
Pull the 7B distill today – takes about 10 minutes including install. Run one real query from your actual workflow, not a benchmark. If the output quality works for that task, you’ve insulated yourself from the Entity List update whenever it lands. If it falls short, you’ve learned something concrete about your quality bar: size up to 32B, or fall back to a US-hosted model. Either way, you’re deciding based on a test you ran, not a headline you read.
FAQ
If DeepSeek gets blacklisted, can I still legally use the weights I already downloaded?
Yes. Entity List restrictions cover exporting US-origin goods, software and technology to the listed entity – not your possession of weights they previously released under an open license. The MIT license you accepted doesn’t get revoked retroactively.
Is running DeepSeek locally actually safe given the security concerns?
The State Department concerns center on DeepSeek the company – alleged military ties, alleged shell-company chip procurement, data sent through their hosted API. A weights file running offline is a different thing entirely. Download the model, disconnect your internet, run a query. If your network monitor stays flat, there’s nothing to trust – the risk is gone by design. For extra assurance, run Ollama inside a container with no outbound network rules set at the firewall level.
Should I just switch to Llama or Qwen instead?
Qwen 2.5 and Llama 3.1 are both excellent – and either avoids the policy risk entirely. But if you’ve already built and tuned prompts around DeepSeek’s reasoning style, switching has a real cost. The local route lets you keep that work. Pick based on your existing investment, not the headline.