Mozilla released its first comprehensive State of Open Source AI report on Friday, and the numbers are striking. Open-weight models now handle the majority of production AI tokens. The capability gap between open and closed models has collapsed to just 3.3%. And the ecosystem around open-source AI has matured into a multi-hundred-billion-dollar commercial market with proven revenue models, publicly traded companies, and enterprise adoption at scale.
The report, authored by Mozilla's CTO Raffi Krikorian and built on data from OpenRouter, the Chatbot Arena, a global developer survey of 1,410 respondents, and financial disclosures from the largest AI companies, paints a picture of an industry at an inflection point. Open weights have won on capability and cost. The battle is now moving up the stack.
Open Weights Close the Capability Gap
The most consequential finding in the report is about parity. The open-versus-closed capability gap on Chatbot Arena has fallen from 8.04% in early 2024 to 3.3% in mid-2026. By August 2024, it had already collapsed to 0.5%. DeepSeek-R1 briefly matched the top US model in February 2025. The gap has reopened slightly as closed reasoning models pushed ahead, but it concentrates in a narrow band: reasoning, long-context retrieval, and agentic tasks. On coding, instruction-following, and general knowledge, open is at or near parity.
The cost curve is even more dramatic. GPT-4-class inference has fallen 50x in 36 months, from $20 per 1 million tokens to roughly $0.40. That is faster than dotcom-era bandwidth or PC-compute price curves. On OpenRouter, the five highest-volume models are all open weights. Chinese open-weight models rose from under 2% of tokens in late 2024 to more than 45% of weekly traffic by April 2026, and roughly 61% among the ten most-used models.
By mid-2026, the top nine models routed roughly 18 trillion weekly tokens for Chinese-built models against about 5.5 trillion for US-built ones. That is a more than 3-to-1 ratio, according to Financial Times analysis cited in the report. Where developers route by cost, they route to open weights.
The Business of Open Source AI Has Gone Mainstream
Open-weight AI is no longer a research curiosity. It is a commercial market at multi-hundred-billion-dollar scale, built by funded companies and run in production by global enterprises. The report catalogs the financial maturity of the ecosystem with data that would have been unthinkable two years ago.
Databricks has crossed a $5.4 billion run-rate and is in pre-IPO territory. Mistral AI scaled 20x to roughly $400 million ARR in twelve months. DeepSeek reached approximately $220 million ARR and recently raised $7.4 billion at a valuation over $50 billion, backed by Tencent, CATL, and the China National AI Fund. Zhipu AI and MiniMax both went public on the Hong Kong Stock Exchange in 2026. Hugging Face, Together AI, Cohere, and Reflection AI each carry disclosed funding in the hundreds of millions to billions.
Five revenue models are proven at scale: hosted inference, enterprise platforms, on-prem licensing, fine-tuning services, and harness tooling. The report also highlights a staggering asymmetry. On OpenRouter from May to September 2025, closed models held roughly 80% of usage but captured approximately 96% of revenue. At roughly 90% capability parity, closed models cost about 6x more per call. The Linux Foundation's Nagle-Yue study estimates the asymmetry amounts to roughly $24.8 billion in unrealized annual savings.
But the report's most sobering finding for closed-model vendors came in June 2026. Three days after Claude Fable 5 went on sale, a single government's export order forced Anthropic to cut access for every foreign national on earth. No other capital was consulted. The models went dark for everyone at 5:21 p.m. on a Friday. As the report puts it: a provider can switch off a model. Nobody can switch off a copy already running on a machine you hold.
The Production Gap Is Still Open's Biggest Vulnerability
For all the good news, the report identifies a critical weak point. Open models lead in adoption: 79% of developers adding AI functionality use them, against 71% for closed, and the two are largely complementary, with 50% of developers using both. But production is where teams stall. Only 51% of open-model teams reach production versus 63% for closed. The gap is operational tooling and trust, not model capability.
Enterprises can buy their way through closed deployment. They do not have the same option for open. The Mozilla stack map scored 48 components across nine layers on 10 maturity criteria. The two coldest columns repeat down every layer: standardization and enterprise readiness. That repeating cold edge is what Mozilla calls the operational gap.
Among developers who churned away from open models, the biggest blockers were infrastructure cost (27%), security and compliance (26%), ongoing maintenance (24%), and deployment complexity (23%). The challenges are operational in every region. In South Asia, security and support concerns spike to 39% and 31% respectively. Only North America and Greater China have more than 15% of developers reporting no major challenges.
The report also flags geopolitical dynamics that founders cannot ignore. More than 70 national AI strategies are live. The largest source of open weights is China, by design. The State Council's AI Plus Initiative and the national Five-Year Plan codify open-source proliferation as a core directive. China's open-weight models out-downloaded the next eight organizations combined in February 2026. DeepSeek reports 26,000-plus enterprise accounts, and 58% of new AI startups in 2025 included it in their stack, even as at least eight jurisdictions restricted the hosted service. The resolution is architectural: enterprises ban the hosted app and adopt the weights anyway, self-hosted or via Western endpoints.
For founders building on AI today, the takeaway is clear. The model layer is becoming a commodity. The value is moving up to what the report calls the agentic harness: the orchestration loop, tools, memory, sandboxes, and permission model above the LLM. That is where production difficulty concentrates, and where the open-versus-closed, owner-versus-renter contest restarts. The companies that win will not be the ones with the best model weights. They will be the ones that solve deployment, operations, and enterprise readiness for an open-weight world.
