December 1, 2025

Executive Summary

Chinese AI startup DeepSeek released two open‑source models. V3.2 and V3.2‑Speciale, that rival or even surpass OpenAI’s GPT‑5 and Google’s Gemini 3.0 Pro. Powered by a novel sparse‑attention mechanism that dramatically slashes inference costs and supports context windows up to 128,000 tokens, the models achieved gold‑medal performance in elite mathematics and coding contests. DeepSeek’s decision to release the models under the permissive MIT license upends the AI industry’s prevailing business model. Entrepreneurs now have access to frontier‑level AI for free, though data‑residency and regulatory concerns remain.

Full Article

If 2024 was the year of closed AI silos, 2025 is shaping up to be the year open source strikes back. Over the weekend, Hangzhou‑based DeepSeek dropped not one but two gargantuan AI models, V3.2 and V3.2‑Speciale, that it claims match or outperform the best that OpenAI and Google have to offer. In case you missed the fireworks, the Speciale variant scored gold medals in the International Mathematical Olympiad, the International Olympiad in Informatics and the ICPC World Finals. That’s like winning the Nobel Prize, the Fields Medal and a coding bootcamp all on the same weekend.

At the heart of DeepSeek’s leap forward is a clever hack called DeepSeek Sparse Attention. Traditional transformers slow to a crawl as sequences get longer; doubling the length usually quadruples the compute. DeepSeek’s “lightning indexer” selectively attends to the most relevant pieces of context, reducing inference costs by roughly 70 %. You can feed the model 128,000 tokens, about a 300‑page book, and still pay less than a dollar per million tokens decoded. For entrepreneurs building document‑analysis tools or summarizing legal tomes, that’s the difference between burning cash and printing money.

The performance numbers are equally eye‑popping. On the American Invitational Mathematics Examination, V3.2‑Speciale scored 96 % compared with GPT‑5’s 94.6 % and Gemini 3.0 Pro’s 95 %. On Harvard‑MIT’s math tournament it delivered 99.2 %, besting Gemini’s 97.5 %. In coding benchmarks, the everyday V3.2 model fixed 73 % of real‑world software bugs. These models aren’t just good at multiple choice; they can plan multi‑day trips with budget constraints, write code across eight languages and conduct research while keeping their train of thought.

Perhaps most disruptive is DeepSeek’s release strategy. Unlike OpenAI and Anthropic, which lock their frontier models behind pricey APIs, DeepSeek posted full model weights, training code and even OpenAI‑compatible wrappers on Hugging Face. Anyone can download, fine‑tune and deploy a 685‑billion‑parameter model without paying a dime. For startups that previously spent fortunes on per‑token API calls, this could be the ultimate cost‑saver. It’s also a shot across the bow of U.S. export controls; the company openly notes that its models run just fine on Chinese‑made chips.

That openness comes with caveats. European regulators have already labeled DeepSeek’s data transfers “unlawful,” and Italy ordered the app blocked earlier this year. U.S. lawmakers talk of banning the service on government devices. Entrepreneurs need to weigh the savings against compliance risks, especially when handling sensitive user data. Yet the broader implication is hard to ignore: if a lean Chinese upstart can match frontier AI with a fraction of the resources, and then give it away, what happens to the incumbents? As one commenter joked, “Rest in peace, ChatGPT.”

For future billionaires, the opportunity is clear. With DeepSeek’s models, you could build an agentic research assistant, a financial analyst that chews through gigabytes of filings, or a travel planner that juggles cost and itinerary. The barrier to entry is no longer compute; it’s creativity and an eye for regulation. Just remember: with great (open) power comes great legal responsibility.

Subscribe to our newsletter to stay up to date with the latest breakthroughs in AI, Business, Technology, and Mindset.