Mozilla: The Current State of Open Source AI

Narrowing Capability Gap, Collapsing Costs

Open-weight models have nearly caught up to closed models in capability. Based on Chatbot Arena, the gap narrowed to 0.5% in August 2024, and stands at 3.3% as of March 2026. In coding, instruction-following, and general knowledge they are equal or superior, but closed models remain ahead only in reasoning, long-context understanding, and agentic tasks.

The cost of GPT-4-level reasoning has dropped 50x over 36 months ($20 → $0.40 per 1M tokens). This is a steeper decline than the bandwidth or PC performance price curves of the dot-com bust era.

Open Source Wins the Token Shift

Based on OpenRouter, a majority of routed tokens have shifted to open-source models. Open-source token share, which was about 1/3 by late 2025, exceeded a majority in mid-2026. The top 5 high-volume models are all open-source.

As of mid-2026, Chinese models process about 18 trillion tokens weekly on OpenRouter, while US models process 5.5 trillion tokens (a 3:1 gap). When developers make routing decisions based on cost, they choose open source.

High Adoption, Low Production Penetration

Mozilla/SlashData 2026 Developer Survey (n=1,410):

• Open-source model adoption: 79% (closed 71%)
• Production deployment rate: open-source 51% vs closed 63%
• Half of developers use both model types together

The cause of the gap isn't lack of capability but absence of operational tooling and reliability issues. Even large enterprises escape closed deployment through funding, but open-source deployment maturity stagnates regardless of company size.

Barriers to Production (in order)

1. High infrastructure/compute costs (27%)
2. Security/compliance (26%)
3. Ongoing maintenance (24%)
4. Deployment/hosting complexity (23%)
5. Lack of specialized support (22%) 6–9. Model evaluation, fine-tuning, integration, documentation (each 17–18%)

By region, only South America and Western Europe show closed-model adoption ahead of open source.

Open Source Stack: Strong Capability, Weak Operations

Evaluating 48 components across 9 layers found that standardization and enterprise readiness score low across all layers. This is the core of the operational gap.

Business Viability Proven

Databricks $5.4B ARR, Mistral $400M (20x annual growth), DeepSeek $220M ARR + $74B funding. Five revenue models confirmed: hosted inference, enterprise platforms, on-prem licensing, fine-tuning services, and harness tools.

OpenRouter data from May–September 2025: closed models account for 80% of usage while generating 96% of revenue. The price ratio is about 6x. Compared to 2024, preference for open source at cost parity is rising dramatically.