Mozilla Foundation released the 2026 edition of its State of Open Source AI report in July, led by Mozilla CTO Raffi Krikorian. The report says the capability gap between open and closed models narrowed from 8.04% in January 2024 to 0.5% in August 2024, then moved back to 3.3% in March 2026 as closed reasoning models advanced.
Mozilla said open models have reached parity in coding and instruction-following tasks, though they still lag in reasoning and long-context work.
Inference costs have dropped sharply as open models take more traffic
The report says GPT-4-level inference pricing fell from $20 per million tokens three years ago to $0.40 now, a 50-fold drop. Mozilla said that pace of decline is faster than the price reductions typically associated with Moore’s Law.
On mainstream inference broker platforms such as OpenRouter, open-weight models increased their traffic share from about one-third at the end of 2025 to more than half by mid-2026. The five highest-traffic models on the platform are all open-weight models. Across the broader developer base, open models posted a 79% adoption rate, ahead of closed models at 71%.
Chinese open-source models gained ground on OpenRouter
Mozilla gave special attention to the speed of expansion in China’s open-source AI ecosystem. It said Chinese models accounted for more than 45% of OpenRouter traffic in April 2026, and 61% of the 10 most-used models on the platform came from China.
Alibaba’s Qwen recorded monthly downloads in February 2026 that exceeded the combined total of the next eight organizations, according to the report. DeepSeek now has more than 26,000 enterprise accounts, and 58% of new AI startups have integrated DeepSeek into their products.
Mozilla linked that rise to China’s “AI+” strategy announced in August 2025 and the country’s “15th Five-Year Plan” in March 2026, both of which elevated open weights to the level of national strategy as a way to offset semiconductor export controls.
Adoption is high, but deployment still trails closed models
The report also highlights what it calls a deployment gap. Open models show a 79% adoption rate, above the 71% recorded for closed models, but only 51% of teams have successfully pushed open models into production. That is below the 63% recorded for closed models.
Mozilla said the bottleneck is not model capability itself. It pointed instead to infrastructure costs, security and compliance requirements, deployment complexity, and ongoing maintenance.
As a strategic recommendation, the report says the industry should secure ownership across layers such as harnesses, memory, and permission systems while open-source models still hold an edge there. It also described the “agent permission model,” or write surface, as the highest-leverage unresolved issue at this stage.