Mozilla has published its first recurring assessment of the open source AI ecosystem, titled "The State of Open Source AI." The report, released in July 2026, argues that open-weight models have moved from a compromise to a practical choice for most developers, while also identifying key vulnerabilities that could tip the balance back toward closed, rented systems.
Open Models Now Dominate Production Traffic
According to the report, open-weight models now account for a majority of production tokens routed through OpenRouter, with the five highest-volume models on the platform all being open. The platform itself now moves 25 trillion tokens per week, five times the volume from late 2025, when open models represented roughly a third of traffic. Hugging Face hosts 2.5 million public models and 13 million users, including a third of the Fortune 500.
Data from the Mozilla and SlashData 2026 developer survey shows that 79% of developers adding AI functionality use open models, compared to 71% for closed models. Half of developers use both, indicating the two categories are largely complementary. However, production deployment remains a bottleneck: only 51% of teams using open models reach production, versus 63% for closed models. The report attributes this gap to operational tooling and trust, not model capability.
Capability Gap Narrows but Persists
The report identifies a 3.3% capability gap between open and closed models, measured across coding, reasoning, and agentic tasks. On the Chatbot Arena leaderboard, the strongest closed model scored 60 while the strongest open model scored 54. A year earlier, the leading open model managed only 22. The gap has collapsed from 8.04% to 0.5% on some benchmarks, but closed systems still lead on the hardest reasoning and multimodal problems.
Mozilla notes that for most workloads, the frontier is not what builders need. Commodity inputs do not hold pricing power, and value moves up to what the report calls the "agentic harness" - the orchestration layer above the model. The report warns that if the token share of open models stalls while the reasoning gap widens, the current momentum could reverse.
Commercial Market at Multi-Hundred-Billion Scale
Open-weight AI has become a commercial market at multi-hundred-billion-dollar scale, built by funded companies and run in production by global enterprises. Databricks crossed a $5.4 billion run-rate with 65% year-over-year growth. Mistral scaled 20 times to approximately $400 million in annual recurring revenue in twelve months, and recently raised at a valuation over $20 billion. DeepSeek reached roughly $220 million in ARR and raised $7.4 billion at a valuation over $50 billion.
The report identifies five proven revenue models at scale: hosted inference, enterprise platforms, on-premises licensing, fine-tuning services, and harness tooling. It contrasts this with the economics of metered pricing, citing examples like Uber engineers billing $500 to $2,000 per month on AI coding tools, leading Uber to cap spending at $1,500 per engineer. Stripe cut its costs by 73% by switching to vLLM, an open inference engine. Microsoft is exploring Azure-hosted DeepSeek and a Copilot Cowork feature.
Governments and Sovereign Capacity
More than 70 national AI strategies are now live. The report argues that the strategic question has shifted from whether to have a national AI policy to which layer of the stack a country can own. The European Commission has proposed an "open source first" rule for public institutions buying AI. Canada has set a national target to lift business adoption from 12% to 60%. France has committed 109 billion euros to AI. The EU AI Act includes general-purpose AI exemptions.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
China's models, particularly Qwen with 942 million downloads, have out-downloaded the next eight organizations combined. Chinese models account for more than 45% of weekly traffic on some platforms, and 61% among the ten most-used models. DeepSeek has 26,000 enterprise accounts. However, eight jurisdictions have restricted DeepSeek. India has deployed 38,231 GPUs with a 10,372 crore rupee outlay and 600 data labs. South Korea committed $71.5 billion, and Saudi Arabia's Humain project is valued at $77 billion with 1.9 gigawatts of compute.
The Agentic Harness as the New Frontier
The report describes the agentic harness as the layer above the model that includes orchestration loops, tools, memory, sandboxes, and permission models. It is where production difficulty concentrates and where the open versus closed contest restarts. LangChain has 126,000 stars and 60% market share. MCP servers and downloads have grown significantly. Databricks released Omnigent. Terminal-Bench 2.0 and 2.1 show a spread between lab-owned and independent scaffolds.
MCP was donated to the Linux Foundation and now has over 10,000 servers and 97 million downloads. The Linux Foundation formed the AAIF for MCP and A2A governance. A2A is in production with platinum members. Only about 21% of organizations have mature agent governance. The report highlights the need for a portable permission specification that does not yet exist. Memory vendors include Mem0, Letta, Zep, and LangMem. Sandbox providers include E2B, Daytona, and Modal. Observability tools include Langfuse, Phoenix, and LangSmith. Auth platforms include WorkOS, Okta, Auth0, Stytch, and Arcade.
Safety, Misuse, and the Risk of Restriction
The report tracks safety and misuse as unsettled issues. It notes how easily safety tuning can be stripped from open weights. The NTIA has adopted a "monitor, don't restrict" approach, but the report warns this could shift after a major misuse event, particularly involving synthetic CSAM or NCII. Authorization failures have been rated CVSS 9.3 to 9.4. The Future of Life AI Safety Index and contractual assurances from the Linux Foundation are noted as ongoing efforts.
The Window Is Open but Closing
Mozilla's CTO Raffi Krikorian wrote an opening letter for the report, drawing parallels to the early web. He described a June 2026 incident where one of the most advanced models went dark everywhere because a government sent a letter, demonstrating that every business renting that model discovered an off switch belonging to someone else.
"We have been here before," Krikorian wrote. "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. We bet on open the first time. Open won. Together, we can do it again."
The report concludes with a test: look at who is seated in the rooms where AI gets decided, and with what status. "The day they seat the people who keep AI open, portable, and widely deployed on equal footing, the shift from renting to owning will have happened. The window is open now. It is closing slowly enough that we can pretend it isn't, and the lease is shorter than it looks."
Related on Neura Market
- AI Tools Directory: Explore open and closed AI tools for development, deployment, and agent orchestration.
- Automation Marketplace: Find solutions for building agentic harnesses, including memory, sandboxes, and observability.
- Open Source AI Models: Browse and compare open-weight models for production use.