On 8 November 2026, the volunteer developers who reverse-engineer Apple Silicon Macs to run Linux told their own users something most open-source projects never have to say out loud. Somebody had been stealing their code before it was even public. They ran it through a large language model to scrub the fingerprints, then prepared to resell it as a commercial product. Asahi Linux's own progress report, dated 7 November 2026 and reported the following day across the trade press, announced the change. The project is moving every repository from the MIT licence to the GNU General Public License, version 2 or later.
Eleven days earlier, on 29 October 2026, a different team made, on the record, the opposite move for the same underlying reason. Mistral released the open weights for Mistral Large 4, a model with roughly one trillion parameters, under a licence the French company itself calls a modified MIT. It kept almost everything permissive and carved out exactly one use: reselling the model as a hosted service without a separate commercial agreement first.
Two projects, one fortnight. The same licence family, pulling in opposite directions.
Neither move is a footnote. Asahi's downstream work underpins a meaningful share of Linux support on Apple's current hardware. A licence change at that scale touches every distribution and every contributor who has built on top of it since the project started. Mistral's open weights are not a demonstration model either. A one-trillion-parameter release with a working self-hosting path is the kind of asset enterprises and governments actually deploy. The model's own documentation recommends a specific inference server and ships an official container image for it. That is why the one use Mistral chose to restrict is the one that would have let someone else capture the commercial value of serving it.
Asahi had chosen MIT deliberately. The project's own account states the reasoning plainly: a permissive licence was meant to invite the BSDs and other free-software operating systems to share code freely in both directions. That choice depended on one assumption holding. Copying by hand or by diff left a trail a human reviewer could follow. Credit and accountability survived the copy even when the licence did not strictly require either.
The project's own description of what broke that assumption is specific. In-development feature code, still sitting in contributors' personal forks, was lifted before it was ever published. It was then rewritten with the help of a large language model to obscure where it came from, with the stated aim of displacing the project and selling Linux on Apple Silicon commercially. Asahi's maintainers reached for a precedent from their own history rather than inventing a new one. WINE made the identical move, from MIT to the Lesser GPL, in 2002, after watching permissively licensed projects treated, in the project's own words, as free real estate. They read the present moment as the same pattern returning with sharper tools.
The mechanism is specific, not symbolic. A permissive licence lets anyone take a copy, modify it, and ship the result without releasing their own changes. The GPL removes that option. Anyone who distributes a modified version has to release its source under the same terms, which a closed commercial product cannot do without giving away the thing it was built to sell. That is why Asahi's maintainers describe the move as closing a gap rather than making a statement. A statement does not require anyone caught laundering the code to publish what they did with it. A licence change does.
Asahi's own announcement does not address a direct counterweight. OpenBSD developer Bryan Steele objected publicly, on the record, calling the relicensing "disingenuous and misguided" and arguing it happened without the consent of every prior contributor. His stake is concrete, not rhetorical. OpenBSD's own Apple Silicon support depends on Asahi's m1n1 bootloader. Its own licensing conventions mean it cannot adopt the GPL-licensed drivers Asahi ships from here on.
Whether a project can relicense MIT-contributed code to the GPL without individually canvassing every past contributor turns on terms that nobody outside Asahi's own contributor process has published. Nothing available this week settles the question either way. It remains an open legal question that a strongly worded blog post does not close, however sound its stated motive sounds to the people reading it.
Eleven days before Asahi's announcement, Mistral had already answered a related question from the opposite direction. Mistral Large 4 runs roughly one trillion total parameters, with about 52 billion active for any given task. It is a mixture-of-experts design; the model's own documentation narrows that further, to 49 billion of pure routed computation before embeddings and output layers are added back in. Mistral shipped it under what the company calls the Ecosystem Open-Weight licence, and describes it in its own words as a modified MIT.
The company's language is accurate as far as it goes. The model is free to use, modify and fine-tune. The one condition is specific and narrow. A business that wants to operate the model itself as a hosted service, reselling access to it rather than building a product on top of it, needs a separate commercial agreement first. That is a restriction on one field of use. By the test the Open Source Definition itself sets, a licence that restricts even one specific commercial use is not an open licence, whatever its own marketing copy calls it.
Put the two side by side and a pattern holds that has nothing to do with code. A permissive licence is a bet that attribution survives copying, because for three decades copying by hand or by diff left a trail a human reviewer could follow. Generative rewriting breaks that bet from Asahi's side. A laundered copy is now indistinguishable from independent work. Nobody can prove the lineage after the fact, so only a stronger copyleft clause makes the laundering legally costly once it is discovered. Reselling a hosted model breaks the same bet from Mistral's side. A model's weights are a compressed asset whose value is realised by serving it, not by reading its source. So the one use that most threatened the business is the one use the licence now excludes before the fact rather than punishing after it.
Both projects are protecting the thing code quietly stopped being on its own: an asset whose value now survives being copied, rewritten and served, long after anyone can say with certainty who actually built it. It is the Bazaar's own argument, run in reverse. Open contribution works only for as long as a contribution can be traced back to someone who can be held to it.
The two licences are also enforced differently. The GPL's force comes from copyright law. Whoever holds copyright in the laundered code can bring an infringement claim once a derivative surfaces, the same mechanism the Software Freedom Conservancy has used against vendors that shipped GPL code without the matching source. Mistral's carve-out has no comparable backstop. It is a term agreed to on download, enforceable as a contract rather than as copyright. That means Mistral has to catch and pursue a breach itself, rather than relying on a public release to surface the evidence for it. A copyleft clause waits for the violation to become visible on its own. A field-of-use carve-out has to be watched.
For anyone who chooses a licence for a living rather than reading about one, the old checklist asked whether a permissive licence would let a competitor build on your work without paying you back. Asahi and Mistral each answered a harder question. When a human reader can no longer tell where a piece of code or a set of weights actually came from, what exactly is the licence still protecting, and does it still protect it?
A copyleft clause that depends on provable lineage is only as strong as the ability to prove lineage. A field-of-use carve-out only works if the excluded use is the one that actually threatens the business. That means naming the use precisely, rather than writing a broad restriction and hoping it holds up later. Read against each other, Asahi's move and Mistral's move are two organisations doing the same homework, reaching different answers because they started from different assets worth protecting. Either way, the licence stops being a one-time decision filed away once the lawyers sign off. It becomes an operational control, checked against how the asset is actually being used. A security control gets checked the same way, against how a system is actually being attacked rather than how it was attacked five years ago.
The Linux Foundation's own State of Commercial Open Source 2026 survey, published the same week at Web Summit in Lisbon, drew on 225 commercial open-source founders and enterprise leaders. It found 57 per cent naming brand and trust as their strongest competitive advantage, ahead of features or community size, in a survey funded by AWS, Red Hat, Futurewei and the Cloud Native Computing Foundation. Asahi is defending that asset with a licence change, and Mistral is defending it with a carve-out. Neither is protecting the code or the weights themselves; both are protecting the ability to say, credibly, where either one came from and who stands behind it. An accredited repository answers the same question differently, checking software before anyone builds on it rather than trusting its own label.
The United States Air Force's Platform One programme runs exactly that kind of check at a scale most open-source maintainers never see. It maintains Iron Bank, an accredited repository of hardened open-source and commercial software images. Each one is checked and signed before any team builds on it, because a container pulled from an unaccredited source cannot otherwise prove its own ancestry to whoever has to answer for it later. Asahi's relicensing and Mistral's carve-out are two private answers to the same question Iron Bank answers institutionally. Who actually put this together, and can whoever uses it next trust the answer without taking it on faith? Neither Asahi nor Mistral supplies defence systems. Each just answered, under pressure, the same question an accreditation process asks before anything open source reaches a system that cannot afford to get the answer wrong.
I have spent thirty years in a field where a licence was a promise that whoever touched your code next would say where it came from. Two projects just concluded, eleven days apart, that generative AI had quietly broken it from opposite ends, and each rewrote its own terms to compensate. If something you built were rewritten tomorrow by a model good enough to hide the seam, would your own licence still protect what you actually care about? Or only the code it was written to protect twenty years ago?
This is one of seven weekly series in The Hamberger Report. Subscribe on LinkedIn and the next one arrives in your feed.
The views expressed in this article are entirely my own, informed by morethan 30 years of professional experience in architecture, security, andtechnology leadership in New Zealand. I write as director of Te PonoLimited; the views are personal and do not represent the position of anyclient, any government agency, or the New Zealand government. My commentaryon legislation and policy is analytical, drawing on publicly availablesources and my professional expertise in architecture, security, and AIgovernance, and it is politically neutral.
About the Author: Andreas Hamberger is a New Zealand-based enterprise architect and technology strategist. Over 30 years, he has moved from compiling kernels on a 486 to leading cloud, cyber, and AI transformation programmes across government, banking, transport, and aviation. He founded Yoper Linux, served as a technology specialist for Novell during the Linux Wars, and is the author of Space Mafia. Free as in Theft is the hidden history of open source software. He can be reached at andreas@thehambergerreport.com.
This article was produced with AI assistance under my direction. Research,drafting and images pass through a pipeline I built and govern: automatedgates for source verification, forbidden language and political neutrality,and my own review before anything is published. The tools include Claude,Gemini and Openart. The frameworks, arguments and editorial judgements aremine and are the same discipline I apply to the AI systems I audit forclients. AI accelerated the work; the thinking, and the responsibility forit, are mine.
[1] Asahi Linux. "Progress Report: Linux 7.3." 7 November 2026 (also widely reported 8 November 2026). https://asahilinux.org/2026/11/progress-report-7-3/
[2] Bryan Steele. Public statement. X (formerly Twitter). 9 November 2026. https://x.com/canadianbryan/status/2119165376764276869
[3] Mistral AI. "Mistral Large 4 (Le Chonk) is now open weight." 29 October 2026. https://mistral.ai/news/mistral-large-4-oss/
[4] Hugging Face. "mistralai/Mistral-Large-4-1T-A52B" model card. 29 October 2026. https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
[5] TechCrunch. "Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity." 29 October 2026. https://techcrunch.com/2026/10/29/mistral-releases-weights-for-its-trillion-parameter-ai-model-targeting-cybersecurity/
[6] Linux Foundation. "New Linux Foundation Research at Web Summit: 72% of Commercial Open Source Companies Report Net Positive AI Impact; Brand and Trust, Not Code, Ranks as Top Competitive Advantage." 10 November 2026. https://www.linuxfoundation.org/press/new-linux-foundation-research-at-web-summit-72-of-commercial-open-source-companies-report-net-positive-ai-impact-brand-and-trust-not-code-ranks-as-top-comp-1793987251744

