DeepSeek stormed the AI landscape earlier this year, unleashing DeepSeek AI models (V1 and R1) onto the world that were on par with ChatGPT offerings from OpenAI, including the most advanced o1 reasoning model. The news that China achieved parity with the best AI models from the US despite bans on high-end AI hardware tanked the stock market. DeepSeek developed its AI models without access to the latest hardware, which it’s banned from purchasing. Instead, the company resorted to software innovations that helped it reduce costs, as well has using some smuggled processors.
DeepSeek is also believed to have distilled ChatGPT versions to train its DeepSeek models, not that it really matters in a world where the likes of OpenAI, Google, Meta, and others have been scraping the web to train their AI on all sorts of data, ignoring copyrights concerns in the process. The bad news didn’t stop there. DeepSeek released its models as open-source, which meant anybody could download and run them locally without worrying about paying DeepSeek for access, or about the privacy and data security implications of using software from China.
Fast-forward to late March, and DeepSeek has quietly launched DeepSeek V3, its next-gen chatbot that might be used to train the DeepSeek R2 reasoning model that should be released in the coming months. However, DeepSeek didn’t make a big deal about the DeepSeek V3 release (yet). As a result, it might not be easy to determine whether you can access it.
Unlike OpenAI, Google, Anthrophic, Meta, and other AI firms that release new AI features on a regular basis, we don’t have any sort of announcement from DeepSeek. The company might have prepared a blog post to announce DeepSeek V3, but that page is not accessible. Also, DeepSeek didn’t make a big deal about the V3 release on social media.
However, the new DeepSeek V3-0324 model is available online. You’ll find it on Hugging Face, assuming you can download 641GB of files and you have the computing power needed to run the full DeepSeek V3 release.
VentureBeat points to OpenRouter as a way to get free API access to the model, including a user-friendly chat interface. This should help you get started with the chatbot immediately.
Then there’s the actual DeepSeek web product, which is available online like ChatGPT. While I’ve repeatedly warned you to tread carefully when it comes to the official DeepSeek apps, considering all your data goes to China and you have no real control over your privacy, it’s likely that the web version of DeepSeek now runs on DeepSeek V3. You might be able to test it there.
What’s certain is that DeepSeek updated its DeepSeek app for iPhone, which now mentions the DeepSeek V3 model:
Powered by the groundbreaking DeepSeek-V3 model with over 600B parameters, this state-of-the-art AI leads global standards and matches top-tier international models across multiple benchmarks. Enjoy faster speeds and comprehensive features designed to answer your questions and enhance your life efficiently.
Even the App Store visual includes a DeepSeek V3 reference, but that one has been there since January when we grabbed the screenshot above. Like the web version, the mobile app will send your data to China. Also, these official DeepSeek variants will likely feature heavy censorship.
That said, DeepSeek V3 should offer improved performance compared to the previous variants, at least according to those who have tried it. Again, there’s no official blog post announcing V3, and we don’t have a new research paper from DeepSeek. The company mentions DeepSeek V3 on its website, but the link goes to Github, where you can download it.
VentureBeat explains that the entire model can run on consumer-grade hardware like the M3 Ultra Mac Studio. That’s a machine that costs $9,500. Most consumers won’t buy one. But it’s still impressive that it can pull it off.
“Tested the new DeepSeek V3 on my internal bench and it has a huge jump in all metrics on all tests. It is now the best non-reasoning model, dethroning Sonnet 3.5,” a researcher said on X.
The implications are huge for what comes next. DeepSeek R2, DeepSeek’s next reasoning mode,l should launch in the coming months, built on top of DeepSeek V3. The blog speculates that DeepSeek R2 might challenge ChatGPT GPT-5 directly, even though we haven’t seen the latter.
OpenAI released GPT-4.5 in preview mode a few weeks ago, but the new model requires a subscription of some sort. OpenAI also teased that GPT-5 would arrive this spring without offering a specific release window.
Before that, ChatGPT unveiled the o3-mini and o3-mini-high reasoning models, which are available to users. It’s unclear whether DeepSeek R2 would compete against those or the full o3 reasoning model.
Whatever the case, DeepSeek V3 seems to be fast and efficient, given what people say online. DeepSeek R2 might not tank the market like its predecessors did, but it should make an impact if DeepSeek can deliver an even better reasoning AI without increasing costs.
Finally, VentureBeat notes that DeepSeek has changed the tone of its AI. DeepSeek V3 is more formal than the previous models, showing a shift to a more technical approach. In practice, the AI might seem less like a human than before, but that’s not necessarily a bad thing as long as it gets the job done.
That said, I expect DeepSeek to reveal more details about DeepSeek V3soone. While we wait, you can start using the model right away via the methods presented above.
