Mozilla just mapped the gap between "open source AI" marketing and actual open source AI, and it turns out most of the industry is lying.
The Summary
- Mozilla's State of Open Source AI report reveals that most models labeled "open source" fail basic openness standards, while philanthropist David Siegel argues governments and corporations should fund genuinely free AI infrastructure
- The semantic battle over "open" matters: if you can't inspect, modify, and redistribute a model's full stack, it's not open source, it's just available
- Companies are racing to call their models "open" while keeping training data, fine-tuning processes, and evaluation sets locked down
The Signal
Mozilla's comprehensive analysis draws a line in the sand that the AI industry has been deliberately blurring. The report establishes clear criteria for what makes AI genuinely open source: accessible model weights, transparent training data, documented architectures, and reproducible pipelines. When they applied these standards to popular "open" models, most failed. Meta's Llama models, widely touted as open alternatives to GPT-4, restrict commercial use and prohibit inspection of training data. Mistral's offerings include similar limitations. Even models from startups positioning themselves as open-first players often gate-keep crucial components.
The timing matters. As Siegel's Fortune piece notes, we're at an inflection point where the foundational infrastructure of AI is being determined. If proprietary models dominate now, the agent economy gets built on rented land. Siegel, whose endowment backs open-source initiatives, argues that governments treating AI development purely as a private-sector competition misunderstands the strategic infrastructure question. You wouldn't let a single company own TCP/IP. Why let a handful of firms control the intelligence layer of Web4?
"The difference between open-weights and open-source determines whether we get digital infrastructure or digital feudalism."
Mozilla's report identifies three categories of models:
- Truly open: Full weights, architecture, data provenance, and training process documented (rare)
- Open-weights: Model available but training data and methods proprietary (common marketing tactic)
- Closed with API access: No meaningful openness, just commercial availability (the majority)
The fourth category, which both Mozilla and Siegel emphasize, doesn't exist yet at scale: comprehensively open foundation models with the compute resources and institutional backing to compete with frontier proprietary systems. EleutherAI and Stability AI have made runs at this. Both hit funding walls. Governments fund AI research. Corporations fund AI products. Nobody's funding AI infrastructure treated as a public good.
The Implication
If you're building agents or automation tools, check the license and documentation of your base model now. The "open source" label is becoming legally contested territory. EU AI Act compliance, liability frameworks, and export controls all hinge on whether your stack is actually inspectable and modifiable. Models that block commercial use or restrict derivatives aren't open source, they're time bombs in your dependency chain.
Watch how governments respond to Siegel's framing. If AI gets treated like roads or electrical grids instead of like smartphones, funding models shift. The agent economy needs base models the same way the internet needed open protocols. Right now, we're building on quicksand labeled as bedrock.