Open-Weight Isn't Open Source: Why the Label Is Broken
Open-weight and open source AI are not the same thing, and the gap between downloadable weights and fully disclosed training data, code, and evals defines a real spectrum rather than a marketing checkbox. Nearly every model marketed as open source today — Llama, DeepSeek, Qwen, Mistral — is open-weight only, while true full openness stays confined to smaller research models. Regulators, especially under the EU AI Act, are now writing their own definition of "open" tied to concrete exemption criteria, which will matter more than any lab's self-description. For PMs, the license terms attached to an "open" model can cap your user base, restrict redistribution, or evaporate your compliance exemption without warning.
Which models in your stack do you describe as "open source" in a pitch deck, a security review, or a compliance memo — and would that claim survive someone actually reading the license? Most teams inherit the label from a blog post or a model card headline, not from the actual terms. That's a liability sitting quietly in your architecture diagram. This week: pull the license for every open-weight model in your production or fine-tuning pipeline and check three things — whether there's a monetization or monthly-active-user cap, whether you're restricted from using outputs to train competing models, and whether you're relying on an "open source" framing for any regulatory exemption claim. If you're operating in or selling into the EU, map that against the AI Act's actual Article 53(2) test, not the vendor's marketing copy. To be fair, most open-weight licenses are workable for the overwhelming majority of product use cases — the restrictions bite at scale, not at a pilot. But if your roadmap assumes you'll grow past that scale, or that today's exemption still applies next year, you're underwriting a bet regulators haven't finished making. Do the license audit before your legal team does it for you, under worse circumstances.
"Open source AI" spans a real spectrum from open-weight (downloadable weights only) to fully open (weights, training data, training code, and eval public); almost no frontier-scale model clears the full bar.
In 2024, the Open Source Initiative finally shipped a formal definition of "open source AI" after two years of fighting about it. By the OSI's own accounting, almost nothing at frontier scale clears the bar it set. Not Llama.
Not DeepSeek. Not Qwen. Not Mistral.
The models everyone calls open source are, by the strictest reading, open-weight — and that distinction is about to matter a lot more than a semantic squabble.
Here's the verdict: "open source AI" is a marketing banner stretched over a real spectrum, and if you're making build-vs-buy calls based on the word "open" in a model card, you're making them on vibes, not terms. The spectrum runs from open-weight (you can download the weights, everything else stays locked) to fully open (weights, training data, training code, and eval all public). Almost every model you've heard of sits at the restrictive end.
That has consequences for redistribution rights, for regulatory exemptions, and for whether you can actually reproduce or audit what you're shipping.
The Spectrum, Laid Out
Most "open" AI coverage flattens four very different commitments into one word. They aren't the same thing, and the gap between them is where the legal and competitive risk lives.
| Tier | Weights | Training data | Training code | Eval/benchmarks | Example |
|---|---|---|---|---|---|
| Closed | No | No | No | Sometimes | Proprietary frontier APIs |
| Open-weight | Yes | No | No | Sometimes | Llama, Mistral Large, DeepSeek, Qwen |
| Source-available, restricted | Yes | Partial | Partial | Sometimes | Models with field-of-use or scale caps |
| Fully open | Yes | Yes | Yes | Yes | OLMo (AI2), Pythia (EleutherAI), Amber (LLM360) |
The models that actually pass the OSI's Open Source AI Definition validation are real, but they're research-scale, not the models your product is likely running in production. The models running in production — Llama-family, DeepSeek, Qwen, Mistral — are open-weight: downloadable, fine-tunable, deployable, and in most cases still governed by a license with teeth. Meta's Llama license, for instance, caps commercial use above a defined monthly-active-user threshold and restricts using Llama outputs to train competing models.
That's not a footnote. That's a license term that can bind your business model.
The Recipe You Never Get
Here's the one analogy worth keeping: an open-weight model is a restaurant handing you the finished dish. You can plate it differently, add your own garnish, even serve it under your own sign. What you don't get is the recipe — the ingredient list, the exact method, the pantry it was sourced from.
You can't reproduce it from scratch, and you can't fully audit what went into it. Fully open models hand you the recipe too: the training data, the training code, the eval harness. That's why almost nobody does it.
The data question, in particular, is where labs go quiet, because disclosing training corpora invites copyright suits and reveals competitive advantage in the same breath.
To be fair to the open-weight labs: releasing weights at all is a real, costly commitment relative to a closed API, and it's the reason startups everywhere can fine-tune a capable model on their own infrastructure without a usage-based bill from a frontier lab. Meta, Alibaba (Qwen), Mistral, and DeepSeek didn't have to publish weights. The fact that they compete on it, rather than only on benchmark scores, has kept the closed labs honest on price and kept smaller teams in the game.
Regulators Are Now Writing the Definition
This is where the debate stops being philosophical. The EU AI Act doesn't define "open source" in its text, but Article 53(2) sets the actual test for the compliance exemption: release under a free and open license, publish the weights along with architecture and usage information, and don't monetize the model. Recital 102 adds that users must be able to freely access, use, modify, and redistribute it.
Crucially, the exemption evaporates entirely for general-purpose models classified as carrying systemic risk — meaning the largest, most capable open-weight releases get the least regulatory benefit from calling themselves open, as reported by legal analyses of the Act's GPAI provisions.
That's a regulator building its own openness ladder, independent of what OSI or any lab prefers to call itself. Expect that ladder to keep climbing. My prediction, falsifiable within the next 12 to 24 months: at least one major regulatory body — most plausibly the EU, through its GPAI Code of Practice review cycle — will tighten the open-source exemption test to explicitly weigh training-data disclosure, not just weight availability and license text.
When that happens, several models marketed today as "open" will lose exemption status they currently claim, and procurement teams that leaned on the open-source label for compliance cover will need a new argument.
The historical parallel is open-source software's own maturation: it took Netscape's 1998 source release and a decade of license litigation before "open source" meant something enforceable rather than aspirational. AI is compressing that cycle. Licensing fights and regulatory definitions are arriving in the same 18-month window that open-weight labs are still racing each other on benchmarks.
The harder question for you isn't whether a model is open. It's whether the openness you're relying on will still hold up legally by the time your product built on it matters.
Get this in your inbox. AI Rundown Daily delivers original briefings every morning — free. Subscribe →
Frequently Asked Questions
Because "open-weight" and "open source" get used interchangeably in casual coverage, and the labs have no incentive to correct the more flattering term. The OSI's formal Open Source AI Definition requires training data, training code, and eval disclosure alongside the weights — a bar that models like Llama, DeepSeek, Qwen, and Mistral don't clear, since none publish their training corpora. That doesn't make open-weight releases worthless; it makes the label imprecise. Treat "open source AI" in a headline as a claim to verify against the actual license, not a settled fact.
Usually yes on infrastructure cost and control, but not automatically on legal risk. Open-weight models let you run inference on your own hardware, avoid per-token API pricing at scale, and fine-tune without a vendor's approval — real savings once you're past pilot volume. What they don't remove is license risk: monthly-active-user caps, restrictions on training competing models, or geographic limits can force a migration later, which is its own cost. Model the license terms into your total cost of ownership, not just the compute bill.
Two distinct risks. First, legal: if your usage grows past a license's user cap or scope, or a regulator tightens the openness bar your compliance argument leaned on, you may need to renegotiate terms or migrate models under time pressure. Second, safety: open weights mean anyone, including bad actors, can strip safety fine-tuning and redeploy the base model, so misuse risk (deepfakes, malware generation, biosecurity-adjacent queries) isn't something you can fully outsource to the model provider the way you can with a closed API. Neither risk is hypothetical — both are already cited by regulators and researchers as reasons for caution.
