Why a 27-Year-Old MIT Researcher Open-Sourced His AlphaFold Competitor

Structure prediction models decide whether a drug candidate is worth synthesizing before anyone spends six months and a few hundred thousand dollars finding out the hard way. That's why the release strategy around these models matters as much as their accuracy scores, and it's why Gabriele Corso's decision to open-source his work is worth paying attention to.

In 2024, Google DeepMind dropped AlphaFold3, a model for predicting how biologically important molecules, proteins, DNA, small-molecule drugs, fold and interact. DeepMind claimed it was the first AI system to beat traditional physics-based methods on accuracy. That's a big deal in computational chemistry, where physics-based simulation has been the gold standard for decades precisely because it's grounded in first principles rather than statistical pattern matching.

The catch: AlphaFold3 only became available to academic researchers, with usage limits, after months of public pressure from the scientific community. No code. No weights. Just an API you could poke at within constraints DeepMind set.

Corso, then a PhD student at MIT, was building a competing model called Boltz-1 at the same time. He didn't love what he was seeing. "It was a reminder of the way the industry was shaping up," he says. So instead of following the API-gated playbook, he released everything: code, model weights, training data sets, benchmarks. All of it, free to modify and redistribute.

The Open-Source Bet

This is the part that's actually interesting from an engineering standpoint. Releasing a structure-prediction model fully open isn't just a licensing choice, it's a statement about who gets to build the next generation of tools on top of your work.

• Boltz-1 launched as a direct, fully open alternative to AlphaFold3, matching its core capability: predicting 3D structure and interactions of proteins and other biomolecules.
• Boltz-2, the successor, extended the architecture and improved on the original's benchmarks.
• BoltzGen flipped the problem around. Instead of predicting structure from sequence, it generates new proteins designed to bind specific molecular targets, think of it as a generative model for molecular design, similar in spirit to how Sora generates video frames, except the output is a binding protein instead of a clip.

All three shipped under open licenses. No API paywall, no usage cap, no waiting for community pressure to force a partial release.

The bet paid off faster than you'd expect. Corso says all of the global top 20 pharmaceutical companies are now using models built on this work, alongside thousands of biotech companies and hundreds of thousands of individual scientists. That's not a niche academic tool, that's infrastructure.

In January, Corso launched Boltz as a public benefit corporation, a structure that legally obligates the company to weigh public benefit alongside shareholder returns. The plan isn't to abandon the open-source approach now that there's a company attached to it. "Everything we do will be accessible," Corso says. "Some will be open source; some will be available by API. We will continue to make them available as widely as possible."

That's a meaningful commitment, and a slightly unusual one. Most companies that start open tend to tighten access once there's revenue on the table. Corso's framing suggests Boltz will keep some models fully open while gating others behind APIs, a hybrid model that's become increasingly common among AI labs trying to balance community goodwill against the need to actually make money.

It's worth situating this in the broader arc of protein structure prediction. AlphaFold2, DeepMind's earlier release, was itself open-sourced and triggered an explosion of downstream research precisely because anyone could run it, fine-tune it, and build on it. AlphaFold3's more restricted rollout broke from that pattern, and the backlash from researchers was loud enough that DeepMind eventually loosened access. Corso's Boltz project reads like a direct response to that friction: rather than lobby for looser restrictions on someone else's model, build the open alternative yourself and let adoption do the talking.

There's also a technical story buried in the adoption numbers. A model only gets used by thousands of biotech companies and hundreds of thousands of scientists if it's actually good, open-source goodwill doesn't compensate for weak benchmarks in a field this competitive. Boltz-1 and its successors had to hold up against physics-based methods and against DeepMind's own numbers to earn that kind of traction. The fact that pharma companies with massive internal R&D budgets, companies that could easily license proprietary tools, are choosing to build on open infrastructure instead says something about both the quality of the model and the value of not being locked into someone else's API terms and rate limits.

Key Takeaways

Corso's work is a useful case study in how release strategy shapes adoption in applied AI research. A technically comparable model, shipped fully open instead of gated behind partial academic access, ended up embedded across the entire top tier of the pharmaceutical industry within roughly a year.

• Boltz-1, Boltz-2, and BoltzGen are all freely available: code, weights, training data, and benchmarks included.
• Adoption now spans all top 20 global pharma companies, thousands of biotechs, and hundreds of hundreds of thousands of scientists, according to Corso.
• The new public benefit corporation, Boltz, plans to mix fully open releases with API-based access rather than closing off the ecosystem entirely.

For practitioners, the practical upshot is straightforward: if you're working in computational chemistry or structural biology and haven't looked at Boltz's stack, it's probably worth a benchmark run against whatever you're currently using. The tools are free, the weights are public, and at this point the community validating them is enormous.