Mozilla Report: Open Source AI Reaches Near-Parity with Closed Models

Mozilla's CTO Raffi Krikorian authored the inaugural 'State of Open Source AI' report (V1.0, July 2026), documenting the rapid rise of open-weight AI models in production environments. Key findings include: the open-vs-closed capability gap on Chatbot Arena narrowed from 8.04% to a low of 0.5% by August 2024, when DeepSeek-R1 briefly matched top US models in February 2025, before reopening to 3.3% by March 2026 as closed reasoning models advanced. Open models are now at or near parity on coding, instruction-following, and general knowledge, with the remaining gap concentrated in reasoning, long-context retrieval, and agentic tasks. GPT-4-equivalent inference costs dropped 50x over 36 months — from $20 to $0.40 per 1M tokens — outpacing historical bandwidth and compute price declines. On OpenRouter, open-weight models grew from a negligible share to a majority of tokens routed by mid-2026, with the five highest-volume models all being open. The report highlights real-world deployments: a Māori broadcaster training speech models for te reo, PwC fine-tuning models for finance clients on private hardware, a Red Cross medical model with clinical trials in Tanzania, and offline cassava disease detection for East African farmers. Mozilla frames the moment as analogous to the browser wars, advocating for competition, interoperability, and vendor independence.