Mozilla's State of Open Source AI Report: The Capability War Is Over. The Deployment War Just Started.

If you have been watching AI for the last three years, you have seen this movie. An open model drops. It ties a closed frontier model on a benchmark. Twitter cheers. Two months later, the closed lab releases something better. Cycle repeats.

This week Mozilla published their first-ever State of Open Source AI report — 950+ developers surveyed, a 48-component stack map, funding data across 20 companies, and token-routing telemetry from OpenRouter. It is the most rigorous open-vs-closed audit anyone has published. And the punchline is not what you would guess from the discourse.

The punchline is: open AI already won the capability war. It is quietly losing the boring one.

The capability gap is basically gone

Here is the number that should stop you scrolling. The open-vs-closed gap on Chatbot Arena, the most widely-tracked capability benchmark, is 3.3% as of March 2026.

That is not 3.3× worse. That is 3.3 percentage points. The open model that scores 79.0 loses to the closed one that scores 79.3. On a bad day the open model wins.

The gap was 8.04% in early 2024. It collapsed to 0.5% by August 2024. DeepSeek-R1 briefly tied the top US model in February 2025. Then closed reasoning models pulled ahead for a year, and open weights clawed back to 3.3%.

Three-point-three. On coding, instruction-following, and general knowledge, open is at parity. The gap concentrates in three places: long-horizon reasoning, multimodal perception, and agentic tasks.

Here is the uncomfortable part. Most production workloads do not need any of those three. The question stopped being "are open models good enough?" and became "what does your actual workload need?"

Inference costs fell 50× in 36 months

This is the other number. A GPT-4-class token cost $20 per million in early 2023. Today it costs about $0.40.

That is not a typo. 50× in three years. Faster than dotcom bandwidth. Faster than the PC-compute curve. Stanford HAI logged 280× drops on GPT-3.5-class capacity over 18 months. MIT clocked 5–10× per year at the frontier, hardware-adjusted.

And open weights are on the cheap side of that curve. On OpenRouter — the closest thing the AI industry has to a real market — the five highest-volume models are all open weights. By mid-2026, Chinese-built open models route roughly 18 trillion weekly tokens against 5.5 trillion for US-built ones. A 3:1 margin.

So who has the revenue? Closed does. A Nagle-Yue study for the Linux Foundation found that on OpenRouter, closed models held about 80% of usage and 96% of revenue. At roughly 90% parity, closed models cost about 6× more per call. That asymmetry compounds to ~$24.8 billion a year in unrealized savings.

Read that twice. Where developers route by capability alone, they route to open. Where they pay, they pay closed. Something is wrong with that picture, and Mozilla's developer survey tells you exactly what.

The real gap is operational

Here is the finding that the entire AI press missed this week and that is actually the point of the report.

79% of developers adding AI functionality use open models. 71% use closed models. Half use both. Adoption is not the problem. Developers clearly trust open weights enough to build on them.

But production is where teams stall. Only 51% of open-model teams reach production. For closed-model teams, it is 63%.

Twelve-point gap. And it is not a capability problem — the same survey proves open models are good enough. It is not a money problem — they are cheaper. It is an operations problem.

Mozilla asked developers what blocks them. The top answers, in order:

1. High infrastructure or compute costs — 27%
2. Security, privacy, or compliance concerns — 26%
3. Ongoing maintenance and updates — 24%
4. Complexity of deployment, hosting, or scaling — 23%
5. Lack of specialized support — 22%

Read the list again. None of these are about the model being dumb. They are about the model being hard to run. Hosting, patching, scaling, securing, monitoring. The work that nobody on Twitter gets clout for.

Mozilla's own stack map shows the same pattern in a different shape. They scored 48 components of the open AI stack across 10 criteria. Capability scores are high. The two coldest columns — repeating down every single layer — are standardization and enterprise readiness. The report calls this "the operational gap" and treats it as the central thesis.

It is the right thesis. Here is why it matters.

Enterprises can buy their way through closed deployment. Open deployment waits on tooling nobody finished.

Mozilla broke out production rate by company size. For closed models, production climbs from 54% (small teams) to 73% (enterprises). For open models, it barely moves — 53% to 57%.

Think about what that means. Big companies adopt closed AI easily. They hire McKinsey, sign a contract with Anthropic or OpenAI, get SOC 2 reports, done. The closed vendor absorbs the operational burden.

Open deployment does not scale that way. A small team that gets an open model into production does so because one engineer is obsessed with it. When the team grows, that engineer's knowledge does not propagate. The model is still technically there. Nobody knows how to redeploy it after a GPU driver update. Six months later it is dark.

This is the open-source AI version of a problem every Linux admin knows. The software is free. The operations are not.

The sovereignty angle nobody wants to talk about

Mozilla's CTO, Raffi Krikorian, opens the report with a comparison to Mozilla's origin story — one company tried to own the front door to the web, an open community rose up, open won. He is making the argument explicitly: open AI is the same fight, twenty-five years later.

The report drops the most concrete example I have seen of why this matters in practice. Three days after Anthropic released Claude Fable 5 earlier this summer, a single government's export order forced Anthropic to cut off access for every foreign national on Earth.

No other government was consulted. None could have been. The models went dark for everyone outside the US at 5:21 p.m. on a Friday.

A provider can switch off a model. Nobody can switch off a copy already running on your own hardware.

That is the quote from the report. It is also the cleanest one-paragraph argument for local-first AI I have read this year.

If you built a product on Fable 5 that Friday morning, your product broke that Friday night, and you had no warning and no recourse. If you had built it on a copy of GLM-5.2 or Kimi K3 running on your own box, it would still be running. Not because open models are smarter. Because they are yours.

More than 70 national AI strategies are now live worldwide, and Mozilla's argument is that the strategic question has shifted — from "should we have an AI policy" to "which layer of the stack can our country actually own." Open weights are the only layer that is ownable. Everything else is a lease.

What the money says

The funding table in the report is the cleanest signal of where serious capital is going.

• Databricks — $5.4B run-rate, pre-IPO
• DeepSeek — ~$220M ARR, raised $7.4B at $50B+ valuation
• Mistral — ~$400M ARR, 20× year-over-year, in talks at €20B
• Moonshot AI (Kimi) — $3.9B raised
• Zhipu AI (GLM) and MiniMax — both went public via Hong Kong IPO in 2026
• Cohere — open-sourced Command A+ in May 2026
• Reflection AI — $2.13B raised
• Together AI — $1.334B raised
• LangChain — $260M raised, 126k+ GitHub stars, 60% developer share in harness tooling

That last row is the interesting one. The biggest open-source funding winner in the tooling layer — not the model layer — is LangChain. The harness. The thing that makes models usable. That is where venture capital is placing its bet on the operational gap.

Corporates agree. Nvidia, Salesforce, AMD, Google, IBM, ASML, Tencent, CATL, and Schwarz Group are all backing the open ecosystem across model, inference, and tooling layers. The smart money has stopped asking whether open wins. It is funding how open gets deployed.

What this means if you actually build things

Three takeaways, if you are shipping software in 2026:

1. Stop benchmarking models and start benchmarking operations. A 3.3% capability gap is in the noise for 95% of workloads. A 12-point production-rate gap is not. If your team is debating which model to use, you are asking the wrong question. The right question is: what does it take to keep a model running in production for two years?

2. The real moat moved up the stack. Mozilla says it plainly: "commodity inputs do not hold pricing power. Value moves up, to the agentic harness." If you are building value, build it in the layer where models plug in and orchestrate — not in the model itself. The model is becoming a utility. The harness is where the business lives.

3. Local-first is no longer a niche preference. It is the only architecture that survives a Friday-afternoon shutdown. A desktop AI assistant that runs open models locally — one that can browse, run terminal commands, read your files, and automate your work without phoning home to a vendor who can be compelled to switch you off — is not a privacy flex. It is operational continuity.

The boring war is the one that matters

The headline you will see repeated this week is "open AI reached parity with closed AI." That headline is true and it undersells the story.

Open weights are not just as good as closed weights. They are cheaper, more sovereign, more heavily adopted by developers, and routing more tokens than closed models by a 3:1 margin. They are winning every dimension except the one that converts adoption into production: the operational layer.

Mozilla's report is, ultimately, a call to build the missing layer. Standardization. Enterprise-grade deployment tooling. The boring stuff that made Linux runnable in a data center, that made Kubernetes a verb, that made Postgres the default. Open AI needs all of that, and almost nobody has finished building it.

That is the opportunity. Not another model. The stuff between the model and the user.

If you are working on that layer — or just trying to get an open model into production without losing your mind — CopperRiver is built exactly for this. A desktop AI assistant that runs open-source models locally, automates real work, and does not ask a vendor's permission to keep running on a Friday night.

The capability war is over. The deployment war just started. Pick a side.