While the EU is busy switching on its new AI labeling rules, China is quietly shipping the next flagship model: DeepSeek V4 Flash. Free in the browser, open-source for self-hosting, and it shot straight to the top of Hugging Face's trending list. If "DeepSeek" still makes you think only of the panic from early 2026, it's worth a second look.
So what can it actually do?
DeepSeek V4 Flash is what's called a Mixture-of-Experts (MoE) model — picture a huge team of specialists where only a handful actually chip in on any given question. According to Hugging Face, the model packs roughly 284 billion parameters, but only about 13 billion are active for any single answer. The result: responses that hold their own against the big models from OpenAI, Google, and Anthropic, while burning noticeably less compute.
The context window is impressive too: one million tokens. Roughly speaking, that's several thick novels the model can "remember" at once while working with you — handy for long documents or entire codebases.
What's free — and what actually costs money?
Chatting on the official DeepSeek website and app is free, no subscription required. Developers also get an API that currently sits in the low-cents-per-million-tokens range — noticeably cheaper than comparable offerings from OpenAI or Anthropic. If you want to save even more, look into "cache hits": asking something the model has already seen recently gets noticeably cheaper.
Then there's the actual surprise: the model weights themselves are fully open under an MIT license. That means anyone can download, modify, and even use them commercially — no legal hoops required. The catch: at 284 billion parameters, this is not something that runs on your laptop. Real self-hosting needs either a small data center or a serious cloud budget. Smaller open models (see our Ollama article) remain the more realistic pick for running things at home.
And where does my data go?
This is where it gets interesting. DeepSeek's own privacy policy states that inputs from chat and API can land on servers in mainland China and may be reused for training unless you opt out. Several countries — including Italy, Australia, Taiwan, and South Korea — have banned DeepSeek from government environments over this. Data protection authorities in France, Ireland, Germany, and a few other EU countries are looking into the service.
That's a real point, not a reason to panic. Before you reflexively write it off: ChatGPT, Gemini, and friends happily collect data for training too — just on servers in the US instead of China. The difference is mostly legal, not necessarily moral. If DeepSeek still gives you an uneasy feeling, do two things: don't type sensitive information into the chat, and check whether one of the smaller open models — the kind you can run locally with Ollama — covers your actual use case. Then everything stays on your own machine.
There's also a practical fix for the server-location worry: some European providers host open models like DeepSeek in EU data centers with a proper data-processing agreement — Cortecs, for instance, had DeepSeek's own V4 Pro model on offer, which we tested ourselves in our big model comparison. The price usually sits a bit above the Chinese original, since a data center in Europe simply costs more than one in China — but if you want it, you can get the same open model's quality without the detour through Chinese servers.
Is this for me?
To try it out: yes, right now. Go to the official DeepSeek site or app, ask it something you actually care about, and compare the answer to what ChatGPT or Claude gives you. Costs nothing, takes two minutes.
As your daily go-to tool: weigh how sensitive your typical inputs are. For harmless questions (recipe ideas, trip planning, a bit of coding help), there's little reason not to. For anything involving personal or professional details, it's worth looking at more privacy-friendly alternatives out of Europe.
For developers, the MIT license is the real headline: a model at this level that you're free to modify and build into your own products simply didn't exist two years ago. That should keep pushing prices down across every provider — good news for anyone using AI, no matter which one you end up choosing.