DeepSeek R2 and why open source AI might win

The pattern is becoming undeniable.

OpenAI releases a frontier model. Three to six months later, an open-weight model matches it. GPT-4 was unmatched for about 9 months. o1’s reasoning advantage lasted about 6. o3’s advantage might last 4.

DeepSeek R2 just matched o3 on reasoning benchmarks. Open weights. Free. Downloadable.

The recurring pattern

Closed model leads. Open model catches up. The gap narrows with each generation.

Meta’s Llama. DeepSeek. Mistral. The open-weight community is deep, well-funded, and getting faster at closing the gap.

Why? Because the fundamental research isn’t secret. The architectures are published. The training techniques are discussed at conferences. The data is similar. The main advantage of closed models is compute budget, and DeepSeek showed that clever engineering can compensate for compute differences.

The equilibrium I’m starting to see

Long-term, I think the equilibrium is: open-weight models at or near parity with closed models on capability. Closed model companies competing not on intelligence but on speed (releasing first), safety (alignment and guardrails), integration (APIs, tools, community), and reliability (uptime, support, enterprise features).

That’s the same equilibrium as Linux versus proprietary operating systems. Linux is free and powerful. Red Hat (now IBM) makes money by selling reliability, support, and enterprise integration on top of it.

Anthropic’s business then isn’t “we have the best model.” It’s “we have the safest, most integrated, most reliable AI platform.” OpenAI’s business is “we ship first and integrate with Microsoft’s platform.”

Model quality becomes table stakes. Everything else becomes the differentiator.

What this means for the AI industry

If open-weight models are at parity, the barriers to building AI products drop. Startups don’t need to pay for API access to a frontier model. They can run Llama or DeepSeek on their own infrastructure. The cost of intelligence approaches zero.

When intelligence is free, the value shifts to what you do with it. The application layer. The user experience. The domain-specific knowledge. The integration with the physical world.

I think this is good for everyone except the companies whose entire business model is “we have the best model and you have to pay us to use it.” That business model has an expiration date, and DeepSeek R2 just moved the date closer.

On Hugging Face, right now, you can download a model that matches the best AI in the world and run it on your own hardware. The implications of that sentence will take years to fully unfold. But they will unfold.

Related thinking: