Mozilla has published the inaugural State of Open Source AI report — V1.0, dated July 2026, recurring — and the CTO’s opening letter leads with a New Zealand example. “In New Zealand’s far north, a Māori broadcaster trains speech models for te reo — a language too small for any market — under a license that keeps the data with its people,” writes CTO Raffi Krikorian. The line is not a tokenistic nod. It is the thesis of the entire report: open weights have closed the capability gap, collapsed the cost of intelligence, and made use cases possible that no closed vendor would ever serve.
🔍 THE BOTTOM LINE
Open-weight models now route a majority of production tokens on OpenRouter. The closed-vs-open capability gap on Chatbot Arena has collapsed from 8.04% in early 2024 to 3.3% in March 2026 — and on coding, instruction-following, and general knowledge, open is at or near parity. The price of GPT-4-class inference fell 50× in 36 months, from $20 to $0.40 per 1M tokens. Mozilla’s headline claim is that the contest has moved one layer up: commodity intelligence no longer holds pricing power, and value is migrating to the agentic harness. The report is also honest about where open is losing — production deployment, standardisation, enterprise readiness — which is what makes it credible.
Open Weights Now Route Most Production Tokens
The most striking data point in the Mozilla report is the OpenRouter token-volume split. By mid-2026, the five highest-volume models on OpenRouter are all open weights: DeepSeek V4 Flash, Xiaomi’s MiMo-V2.5, Tencent’s Hy3 preview, MiniMax M3, and Owl Alpha. Anthropic’s closed Claude models are the next US-built entrants, well behind. The Financial Times analysis Mozilla cites puts the top nine models at roughly 18T weekly tokens for Chinese-built open models against ~5.5T for US-built ones — more than 3:1.
Mozilla is careful with the caveat: by request count, closed US providers still lead. The open lead is a token-volume lead, concentrated in coding and agentic workloads where cost-per-token matters most. But the trajectory is the story. The share of OpenRouter tokens routed through open weights grew from a negligible base to a third by late 2025 to a majority by mid-2026. This is the same curve we tracked in our open-weights LLM convergence analysis — the Mozilla report is the first mainstream confirmation that the crossover has happened.
The Capability Gap: 8.04% → 0.5% → 3.3%
Mozilla’s open-vs-closed gap chart on Chatbot Arena tells a more complicated story than “open is catching up.” The gap collapsed from 8.04% in January 2024 to 0.5% by August 2024. In February 2025, DeepSeek-R1 briefly matched the top US model. Then the gap reopened to 3.3% by March 2026 as closed reasoning models pulled ahead.
Mozilla’s framing is the honest one: 3.3% is an average over a jagged frontier. Open is at or near parity on coding, instruction-following, and general knowledge. The gap concentrates in reasoning, long-context retrieval, and agentic tasks — exactly the workloads where closed labs are spending their frontier budget. The practical implication, Mozilla argues, is that the question is no longer “are open models good enough?” It is “what does your specific workload need?” For most production traffic, the answer is “an open model.”
Inference Fell 50× in 36 Months
The cost curve is the most under-appreciated number in the report. GPT-4-equivalent inference cost fell from roughly $20 per 1M tokens to $0.40 — a 50× drop in 36 months. Mozilla cites Stanford HAI’s AI Index 2025 (a 280× drop for GPT-3.5-class models over 18 months), Epoch AI (9–900× annual decay at the frontier), and a November 2025 MIT study (5–10× per year, hardware-adjusted). The rate of decline is faster than the dotcom-era bandwidth or PC-compute price curves.
This is the structural fact behind the frontier-lab margin collapse we covered in June. Closed models are sold by the token, and at production scale the meter becomes the problem. Open weights, run on hardware you own or lease, have no meter. The Mozilla report names Databricks at a $5.4B run-rate, Mistral at ~$400M ARR (20× in twelve months), and DeepSeek at ~$220M ARR having just raised $7.4B at a valuation over $50B. Five revenue models are proven, Mozilla says: hosted inference, enterprise platforms, on-prem licensing, fine-tuning services, and harness tooling. Open-weight AI is a commercial market at multi-hundred-billion-dollar scale.
Where Open Is Losing: Production, Standardisation, Enterprise
The report is not a victory lap. Mozilla’s own developer survey (with SlashData, 2026, n=1,410) finds 79% of developers adding AI functionality use open models — but only 51% of open-model teams reach production, against 63% for closed. The gap is operational, not capability. Enterprise teams can buy their way through closed deployment; open deployment waits on tooling nobody has finished. The biggest churn drivers for open-model developers are infrastructure/compute costs, security and compliance, maintenance, and deployment complexity — not model quality.
Mozilla’s stack map scores 48 components across 1,361 projects on five maturity criteria. The two coldest columns repeat down every layer: standardisation and enterprise readiness. That repeating cold edge is the operational gap. It is the same gap we noted in our NZ sovereign AI build analysis — the open weights are good enough, the deployment story is not.
The NZ Angle Is the Thesis
Krikorian’s opening paragraph namechecks four other examples alongside Aotearoa: PwC fine-tuning an open model on the language of finance and running it on its own hardware with no per-token meter; Red Cross researchers in Lausanne building an open medical model tuned to humanitarian guidelines, with clinical trials pending in Tanzania; East African farmers diagnosing cassava disease with an on-device model that runs offline in fields the cloud has never reached; and a Swiss public consortium training a national model on public supercomputers and releasing everything. “None of them asked permission, and none of them could have rented this. They own it — that is the whole idea.”
The Māori example is the same one we covered in our Māori data sovereignty and AI voice models piece from May — a Waikato team building a te reo TTS system where the community owns the data, the model, and the output. Mozilla is using it as the lead example because it is the cleanest case for open weights: a language too small for any market, a community that would never be served by a closed vendor’s roadmap, and a licence that keeps the data with its people. The Swiss Apertus sovereign model is the same pattern at national scale. The Mozilla report is implicitly arguing that the sovereign AI story and the open-weights story are the same story — and that NZ’s far north is where both are most legible.
❓ FAQ
Is the open-vs-closed capability gap really only 3.3%? On the Chatbot Arena aggregate, yes, as of March 2026. Mozilla is explicit that this is an average over a jagged frontier: open is at parity on coding and instruction-following, and the gap concentrates in reasoning, long-context retrieval, and agentic tasks. If your workload is agentic reasoning, the gap is wider than 3.3%. If your workload is coding or general knowledge, it is effectively closed.
What does “open weights route a majority of production tokens” actually mean? OpenRouter is a routing layer developers use to call many models through a single API. By mid-2026, more than half of the tokens routed through it go to open-weight models. By request count, closed US providers still lead. The open lead is concentrated in coding and agentic workloads where cost-per-token is the dominant variable. It is a production-volume signal, not a revenue signal — but it is the first mainstream confirmation that the crossover has happened.
Why does Mozilla open with a Māori broadcaster? Because the te reo case is the cleanest argument for open weights: a language too small for any closed vendor’s roadmap, a community that would never be served by a per-token pricing model, and a licence that keeps the data with its people. Mozilla is framing sovereign AI and open weights as the same story. The NZ example is the thesis, not the disclaimer.
What is the catch for New Zealand? The same catch Mozilla names for everyone: deployment. NZ has the use cases (te reo, sovereign AI, small-market specialist models), the open weights are good enough, but the operational tooling — hosting, fine-tuning infrastructure, compliance, maintenance — is not finished. The NZ sovereign AI build is constrained by compute and deployment, not by model capability. Mozilla’s report confirms the constraint is industry-wide, not NZ-specific.
Is Mozilla’s report independent or advocacy? Both, openly. Mozilla is an open-web advocacy organisation and the report is explicitly framed as “a map: where open AI is winning and where it is exposed.” The fact that it names open’s weak points (production deployment, standardisation, enterprise readiness) rather than burying them is what makes it more credible than a vendor whitepaper. The underlying data sources — OpenRouter, Stanford HAI, Epoch AI, MIT, FT — are independent.
🔍 THE BOTTOM LINE
Mozilla’s first State of Open Source AI report is the first mainstream document that says plainly what the data has been showing for a year: open weights route most production tokens, the capability gap has collapsed for most workloads, and the price of intelligence fell 50× in 36 months. The contest has moved one layer up, to the agentic harness, and the operational gap (deployment, standardisation, enterprise readiness) is where open is now losing. The report opens with a Māori broadcaster in NZ’s far north because that case — a language too small for any closed vendor, a community that owns its model — is the cleanest argument for open weights that exists. For New Zealand, the report confirms what the sovereign AI build already showed: the models are good enough, the deployment story is not, and that gap is industry-wide. The open-weight future is not a prediction. It is the current production graph.
