Johansen AI LogoJohansen AI Solutions
All Articles
Open SourceLicensingSovereign AILegal RiskVenture Architecture

Open Weight is Awesome, But… It Is Not Open Source

The legal and ethical nuances of the AI era

Why 'Open Weights' like Llama and Qwen aren't truly 'Open Source' in the OSI sense, and what that means for your startup's long-term risk.

John K. Johansen

We rely heavily on GPT-OSS and Qwen3 in our Sovereign AI Lab. They provide the reasoning power that makes autonomous agents possible without cloud lock-in.

But as a Venture Architect, I have to warn you about a dangerous linguistic shift: "Open Weight" is NOT "Open Source."

As we reach May 2026, the distinction between these two concepts is becoming a major risk factor for startups.

The Semantic Trap

Traditional Open Source (GPL, MIT, Apache 2.0) guarantees your right to use, modify, and distribute the code without restriction.

"Open Weight" models (like Llama 3, Qwen 3, or Mistral) give you access to the result of the training, but they usually come with:

  • Usage Restrictions: Limits on commercial scale (e.g., Llama's 700M user cap).
  • Training Prohibitions: You cannot use the output to train "competing" models.
  • No Data/Code Access: You don't get the training dataset or the specific recipes used to create the weights.

The Risks to Your Startup

1. The Licensing Rug-Pull

If you build your entire AI Agent strategy on a model that can change its license at any time (as we've seen in the broader OSS ecosystem), you are building on quicksand.

2. The IP Ambiguity

Who owns the output of an "Open Weight" model? While current US law says AI-generated content isn't copyrightable, the specific terms of service of the weight provider might attempt to claim rights or impose liabilities.

3. The Lack of Auditability

True Sovereign AI requires transparency. With Open Weights, you are running a "black box" that you cannot fully audit for bias, security vulnerabilities, or copyright infringement in the training data.

The Venture Architect's Path Forward

Does this mean we stop using Qwen or Llama? No. They are too powerful to ignore. But we manage the risk:

  1. Prefer Truly Open Models: We keep a close eye on projects like GPT-OSS and OLMo that aim for full data and code transparency.
  2. Modular Agent Design: We build our Kaigents framework to be model-agnostic. We can swap Qwen for a truly open model in hours, not months.
  3. Document Everything: We maintain a strict SBOM (Software Bill of Materials) that includes the specific license versions of the weights we use.

Conclusion

"Open Weight" is a gift to the engineering community, but it is a commercial dependency. Treat it as a Venture Architect would: enjoy the benefits, but build your moat around the process and the data, not the weights themselves.

Transparency is the only true sovereignty.


John K. Johansen is a lifelong advocate for Open Source and a strategist for the next generation of AI-driven enterprises.

// want to go deeper?

I write about AI agents, startup engineering strategy, and building systems that let small teams do big things — without handing your IP to cloud providers.

Read more articles