By open-sourcing its model, Boltz created a feedback loop where the community discovered novel use-cases, like a crude but effective "inference-time search" for antibody prediction. This demonstrates how open access allows external users to find creative applications the original developers hadn't considered.

The creation of OpenFold was driven by former academics in industry who missed the collaborative models of academia. They saw that replicating DeepMind's restricted AlphaFold tool individually was a massive waste of resources and sought to re-establish a shared, open-source approach for foundational technologies.

OpenFold's strategy isn't just to provide a free tool. By releasing its training code and data, it enables companies to create specialized versions by privately fine-tuning the model on their own proprietary data. This allows firms to maintain a competitive edge while leveraging a shared, open foundation.

Releasing a frontier open-source model successfully is a major operational challenge. It requires tight co-design and coordination between the model lab, hardware vendors, inference engine teams like VLLM, and distribution platforms like Hugging Face to ensure the model is usable and performs well from day one.

Unlike closed-source models where release timing is constrained by inference costs, open-source models benefit from being released "as soon as possible." This strategy helps capture developer loyalty and community engagement, as users run the models on their own infrastructure, freeing the lab to focus on training the next generation.

Fears that universal tools reduce differentiation are misplaced. Instead of just leveling the playing field, open tools like OpenFold raise the entire industry's baseline capability. This shifts competition away from who builds the best foundational model to who can ask the most insightful scientific questions.

Contrary to past momentum, the most advanced AI startups are increasingly adopting and fine-tuning open-source models. This shift is driven by the need for cost-effective speed and deep customization as their workloads mature and scale.

Users on Twitter figured out how to use AlphaFold to predict protein-protein interactions—a key capability the DeepMind team was still developing separately. This highlights the power of open models to unlock emergent capabilities discovered by the community.

The fear that open source will erode the business of OpenAI and Anthropic is misplaced. As open source models make existing solutions cheaper, they compel frontier model providers to tackle the vast number of more complex, unsolved problems, effectively expanding the entire market.