iPhone users turn on to DeepSeek AI

As if from nowhere, OpenAI competitor DeepSeek has somersaulted to the top of the iPhone App Store chart, overtaking ChatGPT’s OpenAI. It’s the latest in a growing line of generative AI (genAI) services and seems to offer some significant advantages, not least its relatively lower development and production costs. You can also ask it how many R’s the word “strawberry” contains and expect an accurate response.

Now on iPhones

Released last week, the DeepSeek app raced to the top of Apple’s App Store charts in multiple countries, including the US. People using the app have noted that the genAI tool can match or beat other similar models in performance.

It also does so at a fraction of the development and deployment costs. It’s also free to use on the web and on the iPhone. In other words, for the price of nothing, you get all the genAI utility you can expect from ChatGPT.

What the industry says

Nvidia’s senior research scientist, Jim Fan, calls DeepSeek “the biggest dark horse” in the open-source LLM field, praising the extent to which the developers have managed to deliver such power with such scant resources.

“We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive — truly open, frontier research that empowers all. It makes no sense. The most entertaining outcome is the most likely,” he wrote on social media.

What’s the market model?

DeepSeek was introduced as open-source models the Chinese developer believes can compete with OpenAI’s and Meta’s best systems. That means the models are available under an MIT license from the popular Hugging Face platform, which itself means these models can be used commercially and without restrictions. Theoretically, even Apple could use it — and many developers are already trying it on relatively modest hardware.

The full package of DeepSeek’s R1 models is available and costs almost 95% less than OpenAI wants for its o1 models. There’s more information available on Github, including an in-depth 30-page technical report.

How good is it?

DeepSeek says its R1 model surpasses OpenAI o1 on the AIME, MATH-500, and SWE-bench Verified benchmarks.  It contains 671 billion parameters, a massive number that means the model can perform very well.

Of course, most on-device AI can’t possibly handle that many parameters, so DeepSeek has made smaller versions of the same model available, the smallest of which should run on an old Mac.

DeepSeek R1 is also built as a self-checking reasoning model, which helps it avoid some of the stupid mistakes other models make. While that reasoning means responses can be a little slower to arrive, they tend to be more reliable. 

Toward an open-source AI

“It shows that open-source AI is catching up,” and in the future we’ll have a multiplicity of such models, rather than just the big commercial models, The Atlantic CEO Nicholas Thompson points out.

One estimate suggests the models might have been trained on a budget as small as $6 million. In comparison, while Meta’s most recent Lama used an estimated 30.8 million GPU-hours to train, DeepSeek required just 2.8 million GPU-hours, according to Andrej Karpathy at EurekaLabs.

In other words, rather than throwing money at a problem, the Chinese researchers are figuring out how to get more from less.

It is impressive that DeepSeek seems to have succeeded in matching OpenAI and Meta’s AI at approximately 10% of the resources, cost, and parameters.

DeepSeek’s researchers said DeepSeek-V3 used Nvidia’s H800 chips for training. (Not everyone accepts the explanation. Scale AI CEO Alexandr Wang expressed doubts about this claim, but still calls the introduction of DeepSeek “earth-shattering”.)

To achieve this, the developers achieved significant technological breakthroughs, such as the capacity to predict consecutive words in a sequence, rather than just the next word. They also figured out to make the system answer questions more efficiently. This is explained well by Thompson.

Good for everyone?

China has figured out how to deliver powerful AI while using fewer resources — and (perhaps most significantly on a planet equipped with finite resources) far less energy.

Is this a bad thing for US interests? Almost certainly not. The fact that China achieved this on such limited resources should be a wake-up call to the US government and investor communities that it’s possible to deliver this technology at much lower costs.

“If there truly has been a breakthrough in the cost to train models from $ 100 million+ to this alleged $6 million number, this is actually very positive for productivity and AI end users, as cost is obviously much lower meaning lower cost of access,” Jon Withaar, a senior portfolio manager at Pictet Asset Management, told Reuters.

That’s a good thing, assuming AI is a good thing in the first place. But it’s a less good option for the big developers in the space. AI stocks are taking a battering today as investors evaluate the achievement. They want value for money, and if DeepSeek can get for $1 what other companies spend a sawbuck on, they’ll want to invest in that.

Ideological AI

It is worth mentioning one other limitation of the system. As it is a Chinese model, it is benchmarked by the Chinese Internet regulator who ensures the genAI responses “embody core socialist values.”

What’s interesting about that is the extent to which this shows how AI models — from China, or from anywhere else — can be built to bake in sets of values that may do more than just reflect their society. No wonder OpenAI wants the US government to invest in US AI.

Getting more for less

If it is indeed correct that DeepSeek has been able to achieve this degree of performance at such low costs using lower-specified tech, it suggests:

That while cash is required to enable the tech, the biggest currency is creative innovation, which flourishes most in open environments. 

That the social and environmental costs in terms of energy, water, and technology we expect AI to require can be dramatically reduced. 

It’s good business to do so.

These reduced costs make AI more accessible to a wider number of developers.

Some of the implications of this are explained in more depth here. But if you’re searching for an iPhone app that manages to capture the technology story while reflecting the evolving global geo-political tension and conversation around environment and industry, you can download it at the App Store today.

You can follow me on social media! Join me on BlueSky,  LinkedIn, Mastodon, and MeWe.