The State of Open Source AI: Parity Reached, Production Gap Remains

Mozilla's State of Open Source AI report (V1.0, July 2026) delivers hard data: open-weight models have closed the capability gap to top closed models, but a production deployment gap persists.

Key findings

• Capability gap: The Chatbot Arena gap dropped from 8.04% to 0.5% by Aug 2024, briefly matched by DeepSeek-R1 in Feb 2025, then reopened to 3.3% by Mar 2026 as closed reasoning models advanced. Open is at or near parity on coding, instruction-following, and general knowledge; the gap concentrates in reasoning, long-context retrieval, and agentic tasks.
• Inference cost collapse: GPT-4-class inference fell 50x in 36 months — from $20 to $0.40 per 1M tokens. That's faster than dotcom-era bandwidth or PC-compute price curves.
• Token volume: Open-weight models now route a majority of production tokens on OpenRouter. The five highest-volume models are all open weights. Chinese-built models route ~18T tokens/week vs ~5.5T for US-built (FT analysis).
• Adoption vs production: 79% of developers adding AI functionality use open models (vs 71% closed), but only 51% of open-model teams reach production (vs 63% for closed). The gap is operational tooling and trust, not model capability.

The report cites concrete use cases: a Māori broadcaster training speech models for te reo under a data-sovereign license; PwC fine-tuning an open model on finance language running on its own hardware; researchers building an open medical model with the Red Cross; farmers diagnosing cassava disease with on-device offline models; and a Swiss public consortium training a national model on public supercomputers, releasing weights, data, and training code.