Mozilla maps the state of open source AI: parity, then a deployment gap

Capability: a gap that keeps resetting

Mozilla has published a new edition of its recurring report on open source AI, mapping where open models stand against the closed frontier. In an introduction, Mozilla CTO Raffi Krikorian frames the project around ownership: speech models built for te reo Māori by a broadcaster in northern New Zealand under a licence that keeps recordings with the communities that supplied them, and a Swiss public consortium that trained a national model on public supercomputers and released the weights, data and training code.

According to the report, the best open models trail the closed leader by about three points on capability indexes while costing roughly 60% of the price, and the distance resets with every release cycle. Mozilla's fit to METR data puts the measured lag between open and closed at about 4.4 months, with open models doubling their capability about every 3.9 months against 5.5 for closed ones.

The frontier is uneven. Open models lead or match closed ones in frontend coding, where one debuted at the top of LMArena's Frontend Code Arena; agentic terminal work is contested, with an open model scoring 88.3 against 88.8 for a closed rival on Terminal-Bench 2.1; professional knowledge work remains a closed edge, where a closed model leads by 92 Elo on GDPval-AA v2. Inference prices for GPT-4-class models have fallen roughly 60-fold in 45 months.

Open weights are not open source

The report is blunt about a vocabulary gap. Sixteen notable open releases were examined, and none ships the data recipe the Open Source Initiative's definition asks for — in most cases “open” means downloadable weights. It also finds that deployment, not capability, holds open models back: they reach production 12 points less often than closed models, a shortfall the report attributes to tooling. Adoption surveys put open models at 79% against 71% for closed ones, with half of teams running both.

Sovereignty, safety and capital

Mozilla argues the case for open weights is optionality. In June, restrictions on foreign-national access left the newest frontier models unavailable to everyone for nineteen days, and the report notes that an order can restrict a hosted API but not weights already downloaded. It cites repatriation pressure — moving a petabyte out of a major cloud provider can cost $90,000–120,000 — and about $24.8 billion in unrealised annual savings from the price gap. Safety tooling is following the path of cybersecurity: the Osprey detection system, used by Bluesky, Discord and Matrix, is being released as shared infrastructure.

Money has moved accordingly. The report lists Nvidia's $12.93 billion purchase of Hugging Face and Stripe's roughly $7.5 billion deal for OpenRouter, alongside $7.4 billion raised by DeepSeek, while noting that 96% of model-layer revenue still went to closed providers in 2025. On the harness layer, the Model Context Protocol has grown from 2 million to more than 110 million monthly SDK downloads in 21 months, with 71,000 public servers.