Mozilla shipped its first State of Open Source AI report this week, and the most interesting parts are not the headline numbers.
Yes, the performance gap between open models and proprietary ones is down to roughly 3.3%—a number that looks nearly closed until you read the fine print, which is that this is an average, and the gap widens considerably on advanced reasoning tasks where closed models still have a real edge. The 50× cost drop over three years (from ~$20 to ~$0.40 per million tokens) is real and structural. These are good starting points. But neither of them is the most interesting finding in the document.
The harness is the product. The report makes a specific claim: the agentic harness—the software layer sitting between people and models, deciding what the AI can see, remember, and do—“can affect performance more than switching the model itself.” They back this with a concrete case: on Terminal-Bench 2.0 (May 2026), a third-party scaffold running Anthropic’s weights reached 79.8% while Claude Code achieved 58.0% on the same weights—a 21.8-point spread favoring the independent harness. Then Anthropic pulled harness development in-house, and by Terminal-Bench 2.1 (July 2026) the gap had collapsed to about 3 points. The lab caught up not by changing the model but by improving the scaffold.
The corollary shows up in vals.ai’s neutral-harness testing: when all models run on the same scaffold, the capability gap between GLM-5.2 (open-weight, MIT license) and Claude Opus 4.8 shrinks to about four points, while the price gap stays at 5×. On a fair harness, open models are roughly competitive; on lab-optimized harnesses, they look weaker than they are.
This has a few implications. First, the model weights are less differentiated than they look. Second, and more strategically uncomfortable for the open ecosystem: whoever controls the harness controls the value. Open weights plus a closed harness is a business model, and several companies are quietly building exactly that. The weights are open; the agent loop is not.
The deployment gap is the other data point worth sitting with. 79% of developers use open models; only 51% actually put them in production. For closed models, that ratio is 63%—meaningfully higher. Mozilla frames this as an “operational tooling and trust” problem rather than a capability problem, which is probably correct. The models are good enough. What’s missing is the surrounding infrastructure for evaluation, deployment, monitoring, and rollback that enterprises need before shipping. This is a solvable problem, but it suggests open source AI is still mostly a developer toy for a meaningful fraction of its current audience.
The geopolitical section is striking in a different way. Chinese models now account for over 45% of weekly traffic on OpenRouter, up from under 2% in late 2024. China’s open-source adoption rate sits at 89%—the highest in the world. The report characterizes this as a deliberate national policy choice rather than organic market behavior, which seems accurate. Treating open weights as sovereignty infrastructure gives governments a reason to route their AI usage away from US providers, and the traffic numbers suggest this is already happening at scale.
The revenue gap—open models driving 1/3 of real-world inference while capturing 4% of the industry’s revenue—is the number that should concern anyone trying to sustain open model development long-term. The labs doing the expensive pretraining runs still need to pay for the compute somehow. The current situation, where downstream applications capture most of the economic value while upstream model development remains expensive, is not obviously stable. Whether this pushes open-weight labs toward hosted services, enterprise support contracts, or something else is an open question. But the 4% figure is a real structural tension in the ecosystem, not just a talking point.
Mozilla’s goal with this report is to establish a “rebel alliance” framing—a coalition of interests opposed to centralized AI control. That framing is more political than technical. But the underlying data is worth reading separately from the advocacy, and several of the findings will reshape how you think about where value actually sits in an AI system.
The answer, increasingly, is not in the weights.