Open-source AI is exploding in creation but concentrated in actual deployment. Stanford's 2026 AI Index reports 5.6 million open-source AI projects on GitHub and Hugging Face, yet across 50 million live domains scanned for real-world usage, the most-deployed open-source AI framework, Botpress, runs on just 1,558 sites compared to 52,682 for OpenAI's closed API. That roughly 34-to-1 ratio reveals the true shape of open-source AI adoption in 2026: massive supply, but concentrated production use.

Why Does the Project Count Explode While Deployments Stay Flat?

The disconnect between 5.6 million projects and real-world adoption tells a story about how AI development has fundamentally changed. When Stanford's count is filtered to projects with at least 10 stars, a low bar for community interest, the 5.6 million shrinks to just 206,880, or 3.7% of the total. Most repositories, Stanford notes, "consist of personal or experimental work and receive minimal attention." On Hugging Face, roughly half of all models have fewer than 200 total downloads, while the top 0.01% account for 49.6% of all downloads.

Stanford

The real driver of the explosion is AI agents themselves. GitHub's Octoverse 2025 report found that about 60% of the year's fastest-growing projects were AI-focused, with more than 1.1 million public repositories now pulling in a large-language-model (LLM) SDK, up 178% year over year. AI agents don't just create new repositories; they commit to existing ones. An analysis of 40.3 million public pull requests from 2022 to 2025 found AI agents now participate in 14.9% of them. Aider's maintainers report that more than 70% of Aider's own code is now written by Aider itself.

Yet these agents run in terminals, CI pipelines, and code editors, not on public websites. That's why they widen the gap between supply and deployment: they accelerate the creation of open-source projects while remaining invisible to domain-level detection.

What Does Open-Source AI Actually Look Like in Production?

The deployment picture is dominated by indie teams and small companies. Of the 1,558 domains running Botpress, 64% of the matched companies have 10 or fewer employees, and 81% have fewer than 50. This is not a story of Fortune 500 enterprises quietly adopting open-source AI at scale; it's a story of small teams choosing self-hosted frameworks for control, cost, and privacy.

Closed APIs, by contrast, dominate the visible web. OpenAI is detectable on 52,682 domains, and MIT Sloan research finds closed models take roughly 80% of all model usage. The reasons are straightforward: closed APIs offer speed to ship, leading-edge performance, and no infrastructure burden. For most organizations, calling an API is faster than managing a self-hosted model.

Hugging Face, the central hub for open-source models, has grown significantly. The platform passed 2 million public models in May 2026, with model uploads more than tripling between 2023 and 2025, reaching 332,000 in a single quarter. Dataset uploads grew fourfold in the same period. Hugging Face now has 13 million users and verified accounts from over 30% of the Fortune 500. Yet these numbers reflect creation and interest, not production deployment.

How to Evaluate Open-Source AI for Your Project

If you're deciding whether to build with open-source or closed AI, the choice depends on your constraints and goals. Here are the key dimensions:

• Control and Privacy: Open-source models run on your infrastructure, keeping data private and giving you full control over updates and customization.
• Cost at Scale: Self-hosting open models can be cheaper than API calls for high-volume workloads, though infrastructure costs are real.
• Speed to Production: Closed APIs get you to market faster with less operational overhead, but you depend on the provider's uptime and pricing.
• Performance Trade-offs: Open models perform at roughly 90% of closed models at release, according to MIT research, but the gap narrows as open models mature.

For most use cases, fine-tuning a pre-trained open model beats training from scratch. The complete path involves defining your problem type, collecting and cleaning data, choosing a framework like PyTorch with Hugging Face's transformers library, building or fine-tuning an architecture, training with proper validation, evaluating honestly, and deploying via API or cloud. Total time typically runs 2 to 8 weeks depending on scope.

Data quality is where most projects actually fail. For fine-tuning a classification model, you need 500 to 5,000 labeled examples per class minimum, with 2,000 or more per class for reliable results. For generative LLMs, 1,000 to 50,000 high-quality prompt-completion pairs produce noticeable behavior change, with quality beating quantity. One consistently undervalued step: spending a day manually reading through 200 to 300 examples of your data to catch label errors and understand edge cases.

What's Driving the Momentum in Open-Source AI?

Despite the deployment gap, the momentum is clearly open. In May 2026, IBM and Red Hat pledged $5 billion to open-source AI. Hugging Face's spring 2026 state-of-open-source report clocked 13 million users and verified accounts from over 30% of the Fortune 500. The organization has become the central hub for model sharing, dataset hosting, and collaborative development.

The real story isn't that open-source AI has failed to reach production; it's that open-source adoption is concentrated, specialized, and growing in specific niches. Small teams building chatbots, classification systems, and custom agents are choosing open models. Large organizations are still reaching for closed APIs first. Both trends are accelerating, and both are real.

For developers and teams evaluating the landscape, the key insight is this: the 5.6 million projects represent potential, not adoption. The 1,558 Botpress deployments represent actual production use. The gap between those numbers is where the real story lives, and it's a story about who controls AI, where it runs, and what trade-offs matter most to the teams building with it.