Open source AI has spent years being treated like the cheaper alternative that might eventually catch up with ChatGPT, Claude, and other proprietary systems.
Mozilla says that argument is now outdated. Its inaugural State of Open Source AI report finds that open models are increasingly competitive, widely used, and far less expensive than they were only a few years ago. Developers clearly want them. Getting those models into production is where things start to fall apart.
Mozilla says 79 percent of surveyed developers currently use open models. Only 51 percent of those users have deployed them in production, compared with 63 percent for closed models. The survey was conducted by SlashData and included 1,494 qualified respondents, although some findings rely on smaller subsets of that group.
The problem does not appear to be a lack of interest.
Developers pointed to infrastructure and compute costs, security and compliance concerns, maintenance, deployment complexity, and integration problems. An open model may be available to download, but that does not mean a company can easily run it at scale.
The gap also remains at larger organizations. Mozilla found that 57 percent of enterprise open model adopters reached production, compared with 73 percent for closed models. Bigger companies can spend their way through many proprietary deployment problems. Open deployments still depend on tools and support that the wider ecosystem has not finished building.
Mozilla puts it bluntly in the report: “The gap is operational tooling and trust, not model capability.”
Open models no longer need to prove that they can generate text, write code, or answer questions. The harder challenge is making them reliable enough for companies that expect documentation, security controls, compliance features, monitoring, and someone to call when something breaks.
Mozilla says the average capability gap between leading open and closed models has narrowed to 3.3 points. That number sounds impressive, but it should not be treated as though every open model is now nearly identical to every proprietary one.
Open models reportedly perform at or near parity in coding, instruction following, and general knowledge. Closed systems still hold an advantage in advanced reasoning, long-context retrieval, and agentic tasks.
A coding assistant and an autonomous agent do not place the same demands on a model. One may work perfectly well with an open alternative, while the other may still benefit from a closed system.
Cost is helping drive adoption.
Mozilla says inference at roughly GPT-4-class performance fell from about $20 to $0.40 per million tokens over 36 months. The report credits open weight releases, including Llama and DeepSeek models, with putting pressure on prices.
Open models now account for roughly one-third of tokens routed through OpenRouter. By June 2026, the five largest models on the platform by monthly token volume were reportedly all open weight models. Mozilla notes that OpenRouter only reflects routed traffic and does not include first-party usage from services such as ChatGPT and Gemini.
The money is not following the usage at the same rate.
Mozilla cites research showing open models generating about 20 percent of model-layer usage during one measured period while capturing only around 4 percent of the revenue. Open systems can deliver substantial value without much of that money flowing back to the organizations maintaining them.
That could become a real problem. Open source AI still needs funding for documentation, security work, deployment tools, evaluation systems, and enterprise support. Downloads and GitHub stars do not pay those bills.
Mozilla believes the next fight will happen above the model itself.
The report focuses heavily on what it calls the agentic harness, the software layer that controls an AI system’s tools, memory, permissions, execution environment, and interactions with users.
That surrounding software can affect performance as much as the model in some situations. It can also create a new kind of lock-in. A company may technically be able to replace one model with another, but switching becomes much harder when its workflows, memory, permissions, and business data are tied to one provider’s platform.
Security is another weak spot.
Closed platforms often include compliance features and safeguards by default. Companies running open models usually have to configure more of that themselves. At the same time, keeping model weights secret does not automatically prevent prompt injection, data leaks, or poorly controlled agent actions.
Permission management may be the biggest unanswered question. AI agents can read information, send messages, alter records, run code, and potentially spend money. Mozilla says the industry still lacks a portable standard defining which actions an agent should be allowed to take without further approval.
That issue will only get worse as businesses give AI systems access to more tools.
Mozilla clearly favors the open source side of this debate, and some of the report should be read as advocacy. The 3.3-point capability gap is an average that hides meaningful weaknesses, while several parts of the report are based on directional assessments rather than hard measurements.
Still, the production numbers are difficult to ignore.
Developers appear ready to use open models. The ecosystem around those models is not making deployment easy enough.
Open source AI has largely proven that it can compete. Now it has to prove that companies can run it reliably without building half the infrastructure themselves.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz