OSI drops official open source AI definition (and Llama doesn't count)

For the past two years, tech companies have slapped the "open source" label on just about any AI model they dropped on the internet, even when key ingredients remained strictly locked behind closed doors.

Now, the referees have finally blown the whistle. On October 28, the Open Source Initiative released version 1.0 of its Open Source AI Definition, setting the first official standard for what genuinely qualifies as open source AI.

Under the new framework, an AI system must offer enough technical detail for someone to substantially recreate it. Developers must provide the complete training code, data processing pipelines, and model weights under approved terms, granting anyone the freedom to modify and build on top of the model for any purpose without special permission.

Crucially, the definition does not mandate that creators upload their full, raw training datasets. Instead, it requires detailed disclosures on data provenance, processing methodologies, and licensing details.

Big Tech gets called out

If you apply these rules today, the current landscape of so-called open models shrinks dramatically.

Meta's Llama is the biggest casualty. Despite Mark Zuckerberg marketing Llama as an open gift to humanity, its license forces companies with over 700 million monthly active users to seek bespoke approval and bans certain commercial use cases. OSI leadership made it clear: Llama is not open source.

According to the OSI, Google and Microsoft agreed to back off from using the term for models that fall short after discussions, while Meta pushed back. Meta argued that AI is too complex for traditional definitions and that its guardrails exist for public safety.

Meanwhile, open-source purists argue the OSI actually set the bar too low by letting companies withhold raw training data. Lightning AI CTO Luca Antiga warned that without clear data licensing, businesses still face major legal blind spots.

Why it matters

We think this definition marks a long-overdue reality check for AI hype. The OSI might not have regulatory police power, but it carries immense cultural and legal weight.

For builders and startups, it cuts through deceptive open-washing. Up until now, teams integrating free weights into their stack often discovered restrictive enterprise traps only after scaling. Having a clear standard helps engineers distinguish between true digital commons and proprietary freemium marketing.

For the broader industry, this battle is really about government regulation. Policymakers in regions like the European Union are creating compliance carve-outs specifically for open-source systems. If regulators adopt the OSI's benchmark, giants like Meta won't be able to enjoy the legal perks of openness while withholding the keys to their data engines.

Looking at the compromises, the decision to not mandate raw datasets feels like a tactical concession to reality, given rampant copyright lawsuits. It is not perfect, but establishing an explicit baseline is the only way to stop Big Tech from moving the goalposts.