Open-source AI models now trail frontier closed models by just 3.3 percentage points on leading benchmarks, yet capture only 4% of global AI market revenue despite powering roughly one-third of all active AI applications, according to Mozilla’s inaugural State of Open Source AI report, published July 14, 2026.
Raffi Krikorian, CTO of Mozilla, authored the report, which surveyed more than 950 developers globally. The report identifies a structural imbalance: 79% of developers use open-weight models, but only 51% have deployed them in production — compared with 63% for closed models. The barrier is not capability, the report finds, but operational complexity: infrastructure setup, fine-tuning pipelines, security audits, and maintaining model versions are harder with open weights than calling a commercial API. AI model costs have fallen up to 50x over the past three years, narrowing the economic case for closed models in cost-sensitive applications.
Why Only 51% of Developers Ship Open Models
The deployment gap — 79% usage versus 51% in production — is the report’s pivotal finding for businesses evaluating open versus closed AI. Mozilla’s data shows the gap is driven by the “agentic harness” problem: the software layer between users and models that controls what an AI system can see, remember, and act on. According to the report, “changing the surrounding software can affect performance more than switching the model itself.” Closed models come with managed APIs that handle this layer; open models require teams to build and maintain it.
Closed models retain meaningful leads in three specific capability categories: reasoning, long-context retrieval, and complex agentic tasks. Outside those categories, the 3.3-point gap has become small enough that organizations running straightforward inference workloads — summarization, classification, structured extraction — can often substitute open models without a measurable quality difference.
Who Is Making Money on Open Source AI
Despite the revenue gap, a cluster of commercial open-source AI companies is growing fast. Databricks is running at a $5.4 billion annual revenue run-rate. Mistral’s ARR reached approximately $400 million, a 20-fold increase in 12 months. DeepSeek reported approximately $220 million in ARR and recently raised $7.4 billion at a valuation above $50 billion; its V4 model is processing 18.4 trillion tokens per month, according to Mozilla’s report.
Regional adoption patterns show sharp divergence. China and East Asia report 89% open-source AI adoption — the highest of any region — with China treating open-source as explicit national strategy. The report frames the overall dynamic in terms Mozilla knows well: Krikorian wrote, “We have been here before. Mozilla exists because one company tried to own the front door to the web, and an open community rose up to make sure it never could. Twenty-five years later, someone is running the same play.”
What This Means for Businesses Choosing AI Tools
For business decision-makers comparing AI tools for business workflows, the 3.3-point capability gap means the evaluation question has shifted. The question is no longer “are open models good enough?” but “does my team have the operational capacity to run them?” Teams with strong engineering resources can increasingly use open models — including Apache-licensed models such as Tencent’s Hy3 — at substantially lower per-token cost. Teams without that infrastructure will continue to pay the closed-model premium for the managed API experience.
The 4% revenue figure is a market structural lag, not a ceiling. As tooling matures to close the operational gap, the revenue share for open models is expected to rise — which makes the current environment a transitional window for businesses that invest now in open-model infrastructure.
For context: The competitive dynamics between open and closed AI models are playing out in real product decisions in 2026. Our coverage of Tencent Hy3 — a 295B open-source MoE model released under Apache license — shows how the open-model frontier has advanced since early 2026.
Our Take: Open models are winning the developer vote but losing the money race — and that gap will close unevenly. The 3.3% capability margin is now small enough to justify a serious evaluation of open models for cost-sensitive applications. The remaining moat for closed models is the managed agentic harness, not raw intelligence. Businesses that build the operational infrastructure now will have significantly lower inference costs when the harness tooling catches up — likely within 12–18 months if the trajectory in the Mozilla data continues.
