Why Open Models Are the Only Sustainable Way to Teach AI

The curriculum gap nobody is talking about: AI consumer vs AI builder.

More and more universities are adding AI to their curriculums: a new module, a workshop on prompt engineering, maybe a course on the latest tools. It’s encouraging to see academic institutions taking action and helping students stay on track with new technology. But it would be even more encouraging if those same institutions made sure students were being taught to understand innovation, not just trust the output of a new tool. There’s a difference between using AI and building with AI, and right now, a big majority of institutions are only teaching the first one.

Using AI means working with what already exists: prompting ChatGPT, plugging into a closed API, integrating a tool someone else built. That has certain value for today’s workforce, but it doesn’t give you any real understanding of what’s happening underneath. AI adopters rely on decisions made by someone else, on a timeline determined by someone else, for a price that can change whenever they choose. Building with AI is different. It means taking an open model, putting your hands around it, applying it to your own data, deploying it, and understanding why it behaves the way it does. It means being able to create and use the best AI for your specific use case, not just consume outputs. That's the difference, in the long run, between a career that gets disrupted and one that keeps evolving with technology.

At Hugging Face, we believe this is the only way anyone can implement AI sustainably.

Why open source is not only a technical choice, but an educational one too.

Open source AI means you can see what you're building with. You can inspect the model, learn how it's architected, audit its performance, and customise it for your exact use case; whether you're in healthcare, finance, law, education, or anything else. You're not captive to a vendor's roadmap or pricing. You're not handing your data to a third party every time you run an inference. You own your stack. A student who has worked this way understands something a student who's only ever touched a closed API doesn't: AI isn't magic. It's engineered. It can be audited, shaped, and challenged.

That shift in understanding, from user to builder, is what Stanford’s Victor Lee has been arguing for: students need to develop the perspective of developers, not just users and critics. The research case is just as clear: a 2025 paper in AI Magazine demonstrated that experiments conducted on closed models, where weights and training data aren’t accessible, are fundamentally not reproducible. This isn’t a niche concern, it’s a problem that undermines the basic standards of scientific integrity, and it’s one the field has been too slow to take seriously. Open models aren’t just a pedagogical preference, but rather a prerequisite for doing research that holds up, and sustains in the long run.

The most forward-looking institutions have already drawn that conclusion. In January 2026, Stanford HAI, ETH Zurich, and EPFL formalised a transatlantic partnership at Davos to develop open-source foundation models that prioritise societal values over commercial interests; explicitly framing it as academia’s response to an industry increasingly monopolised by a handful of closed players. And they’re not alone, regulators have drawn the same conclusion: the EU AI Act, the most significant AI legislation in the world right now, was written with open, auditable models as a protected category; recognising that transparency and reproducibility are public goods, not just technical preferences. Hugging Face co-authored the official developer guidance on its open-source exemptions alongside the Mozilla Foundation and the Linux Foundation. The direction of travel is clear, and the market numbers back it up. The performance gap between leading open and closed models has narrowed from 27 weeks in early 2024 to 13 weeks by mid-2025. Switching to open alternatives could save the AI industry roughly $25 billion annually, not because open models are a compromise, but because they now match or exceed closed performance in most benchmarks.

The employers hiring your graduates are raising the bar.

The World Economic Forum’s Future of Jobs Report 2025 identified AI Model & Application Development as the single hardest skill to find globally. What does this look like concretely? Microsoft laid off 6,000 employees in May 2025 while simultaneously expanding its AI engineering teams. The total headcount ended roughly flat, but the composition shifted: fewer engineers maintaining legacy products, more engineers building AI-native systems. Google, Meta, and Amazon followed the same pattern. AI engineer roles now carry a 20–35% salary premium over equivalent software roles at the same companies. Google went further in 2026, hiring forward-deployed engineers whose entire job is to embed inside client organisations and build working AI systems in production; because the bottleneck, they concluded, is no longer the model, it’s the last mile of integration, and that last mile requires people who can build.

The institutions that close this gap first will gain a reputation that compounds, attracting better students, stronger research partnerships, more industry funding. The ones that don’t adapt face the reverse: graduates who take longer to hire, research that hits reproducibility ceilings, and funding bodies increasingly aligned with open and auditable AI that start looking elsewhere. As the World Economic Forum put it plainly: “We cannot continue to prepare students for jobs that may no longer exist.”

Open source AI has a reputation for complexity it no longer deserves.

Most institutions willing to make this shift often don’t know where to start, lacking a structured path. The open source AI ecosystem is so vast that it seems complicated to penetrate it, and that’s why we built the Academia Hub. Academia Hub is an institution-wide subscription built for everyone working with AI in academia: students learning to build for the first time, and researchers pushing the frontier. It gives institutions access to the open model ecosystem, compute resources, and collaborative infrastructure to do real work: whether that's a student fine-tuning their first model and deploying it as a portfolio piece, or a research lab running reproducible experiments on open models they fully control.

The common concerns on open source AI have real answers on Hugging Face:

• On quality: the Hugging Face Open LLM Leaderboard benchmarks hundreds of open models across standardised tests, giving students and faculty reproducible, comparable scores rather than vendor marketing.
• On security: every model on the Hub is scanned automatically through our partnership with JFrog, eliminating over 96% of false positives while catching real threats; including zero-day models other scanners missed.
• On licensing: the OpenRAIL framework embeds responsible use conditions directly into model licenses, and every model on the Hub displays its license clearly, filterable by type.

At Hugging Face we believe we’re at a moment where the choices institutions make about AI education will shape the next generation of practitioners: who builds AI, how they build it, and whether they build it responsibly and sustainably. The largest employers in the world are restructuring their workforces around a single question: who can build? We want universities to be part of that answer, not as passive adopters of tools built elsewhere, but as active shapers of what open, responsible AI education looks like. The institutions that make that bet now will produce the graduates who define this field. Open source is the secure and sustainable way to adopt AI, for your students, for your research and for your institution’s future.

Explore Academia Hub