Open Weights Tell You Where, Not Whose
Vectara rebuilt its Hughes Hallucination Evaluation Model leaderboard, HHEM, against a harder test this cycle: 7,700 articles dated November 2025 to May 2026. The old table used a smaller archived set this series has quoted for months. The lowest hallucination rate on both tables belongs to finix_s1_32b, a 32 billion parameter model published under the organisation name antgroup, Ant Group [1]. It scores 1.8 per cent on the new document set and 0.6 per cent on the archived original. Factual consistency sits at 98.2 per cent, the answer rate at 99.5 per cent. It tops roughly 105 evaluated models, which span a thirteen and a half fold range in error rate [1][2]. A model small enough to run on hardware an organisation owns outright now holds the best published score on the task a sovereignty-minded organisation cares about most. Staying faithful to its own documents, not a vendor's.
I have spent three decades telling New Zealand boards that self-hosting is a sovereignty decision, not an information technology preference. This is the clearest proof of that argument I have seen. It is also the moment I have to admit the argument was only half the story. The model proving it is published by a Chinese financial technology conglomerate. The question that matters most now is not where the inference happens.
Linus Torvalds posted the first version of what became Linux to Usenet in August 1991. The General Public Licence, the GPL, under which he released it did something precise, not broad. It settled who could use a piece of code, who could modify it, and who could redistribute it. This series has called it a constitution in its own right. Linux 1.0 shipped under that licence in March 1994. Every patch since has carried the same obligation forward: modify it, and the change flows back to the community that can see it. Nobody could fork the kernel into a private branch and keep it. That is a real and durable form of freedom. This series has made that argument for thirty-two years.
What the GPL never promised was an answer to a second question: who wrote the code, and why. For most of Linux's history that gap did not matter in practice, because the answer sat in plain sight. Every commit carries a name. Every patch sits on a public mailing list, argued over in the open, attributed to a person who can be asked what they meant by it. The provenance of a kernel patch was never a separate question from the patch itself. The commit history was the provenance.
An open-weight model inverts that arrangement, in a way this series has not had to reckon with before. The weights carry the GPL's own promise, translated to a different kind of artefact. They are inspectable, redistributable, and runnable on hardware you own outright. Nobody on the other end of a network connection decides whether your request goes through today. What they do not carry is a commit history. A model's weights are the output of a training run. The data it read, and the choices that shaped what it rewards, are not written into the file the way a GPL header is written into source code. You can read every line of what a model produces. You cannot read why it produces it from the weights alone.
A New Zealand organisation that self-hosted finix_s1_32b, specifically to keep its own documents off someone else's servers, would have solved exactly one of those two problems. So I went looking for the answer to the other one: under what licence is this model actually published. It should have been a five-minute check. Instead it took most of an afternoon and ended without an answer.
Ant Group runs two known public organisations on Hugging Face, the platform where open models are normally hosted. The first, antgroup, lists two models, a reasoning assistant and a database query tool, neither named Finix [3]. The second, inclusionAI, describes itself as Ant Group's own artificial general intelligence initiative and states outright that every project it hosts is released under the MIT or Apache licence [4]. It lists nine models. None of them is finix_s1_32b either.
A guessed address at the obvious path returned something more specific than silence: an HTTP 401 Unauthorized response, not the HTTP 404 Not Found a missing page would return [3]. On Hugging Face that pattern usually means a repository exists and is gated, rather than nothing being there. It does not confirm which. A search of the platform for the model's own name returned zero results. A wider search across the open web turned up two models that are not the one on the leaderboard: Finix-P1-32B, a different Ant Group model listed elsewhere as proprietary, and a separate, unrelated "Finix-S1" benchmarked in insurance industry evaluation material. Neither is the model that tops Vectara's table.
None of this proves the model was never released. It proves that finding where it was released took five separate checks. The result: two near misses and one locked door. This is supposed to be exactly the kind of model an organisation can inspect before trusting it with its own paperwork.
Even where the law has tried to settle this question, it has settled it the same way the GPL did: narrowly, and on the side of availability rather than origin. The European Union's AI Act exempts a free and open-source general-purpose model provider from two specific obligations: detailed technical documentation, and an authorised representative inside the Union. That exemption applies where the weights, the parameters, and the architecture are freely available, with no payment required beyond standard attribution and same-licence terms [5]. What it does not appear to reach is a separate obligation: a summary of the training content, and the copyright policy that governed how that content was used [5]. The two obligations sit in different parts of the same Act. Qualifying for one does not excuse the other.
A government that writes rules specifically for artificial intelligence still treats "the weights are open" and "the training is disclosed" as two separate questions, each satisfiable on its own. Nobody had to argue for that distinction. The law simply assumes it. The GPL made the same assumption: that modification and attribution were two different duties, decades before anyone needed it written down for a different kind of file.
For a board or a technical team weighing a self-hosted model against a vendor's hosted one: this is not a reason to avoid open-weight models. The sovereignty case for self-hosting remains exactly as strong as the Vectara numbers suggest. The decision requires two separate questions, not one.
The first question is the one most procurement processes already ask, because it is the one most vendors are happy to answer: where does inference happen, and under whose technical control. Self-hosting finix_s1_32b on your own hardware answers that question completely. Nobody outside the building reads your documents to produce the summary. Nobody's uptime, pricing change, or policy shift reaches into your workflow uninvited.
The second question is who shaped the training run, on what data, and for what reason. Self-hosting does nothing for that question. The model you host is the output of a process you cannot inspect from the weights alone. Answering the first question completely and the second not at all solves half a sovereignty problem, not the whole of it.
Ask both questions before a model goes into production, not after. Ask the first the way procurement already knows how: where does it run, who administers it, what happens if the vendor relationship ends. Ask the second a different way, because the normal channels will not answer it. Can the publisher point to a model card, a training data summary, or a documented licence? Or does a five-minute check turn into an afternoon, the way it did here? An honest "we could not confirm it" is a safer procurement answer than an unexamined assumption that open weights settled the question on their own.
The open-source dimension of this one is the article itself, not a bridge to it. Every fact above came from inspecting public material: Ant Group's own Hugging Face organisations, each stating its own licensing policy in its own words, and a guessed web address that answered with a specific, checkable response rather than silence. inclusionAI's blanket MIT and Apache policy is itself a form of the same transparency the GPL established for code, applied to models that actually carry it through. finix_s1_32b's absence from both organisations makes that absence visible and nameable, rather than buried inside a vendor's closed infrastructure where nobody could have gone looking at all.
Sovereign forks of Linux show what happens when a state treats both questions, not only the first, as its own to answer. Kylin in China and Astra Linux in Russia are not merely rebadged distributions. Each runs its own security review process, its own patch cadence, and its own certification regime. These are built on the inspectable freedom the GPL guarantees, extended to cover a layer finix_s1_32b's publisher has not: who decided what goes in, on what standard, answerable to which authority. Neither state has resolved the deeper question any more fully than Ant Group has. A sovereign fork answers who approved this, not why it was trained this way. The shape of the question is the one a defence procurement officer already asks about commercial software. An open-weight model asks it again, for an artefact national software review standards were not built to read.
For a team running Ubuntu, Kylin, or any other Linux-derived stack in production, the same discipline applies the moment an open-weight model joins the toolchain. Treat the model card the way you already treat a software bill of materials: something you ask for before deployment, not something you discover is missing during an audit. Where no model card exists, say so in your own procurement record rather than letting the gap travel unrecorded into production.
Open Source Summit Europe ran in Prague the week before this piece went to print, under a banner marking thirty-five years since that first Usenet post. Linus Torvalds and Ericsson's Dirk Hohndel shared a stage there on vendor-neutral collaboration [6]. The Linux Foundation's own framing for the week put open source at the centre of a nation's technological independence. Not as a side benefit of writing good code, but as the point of writing it in the open at all [6]. Prague did not resolve the provenance question this piece has been chasing. Nobody there claimed it had. A conference built entirely on the premise that owning your own toolchain matters is exactly the room where the training-provenance question belongs next. It is a room this series has been arguing its way toward for thirty-five years, without ever quite naming it.
In 2001 you bought the linux.co.nz domain and a LINUX number plate at auction. A few thousand dollars, on the belief that owning the name would pay off. By your own account, it has. Self-hosting finix_s1_32b is a different kind of bet on ownership. You own the inference, the same way you owned the domain. But you have no equivalent purchase on how the thing you are running was actually made. Say your own organisation had to choose the model with the best proven honesty score on the market today. You could run it entirely on hardware you control. But you could not get a straight answer out of its publisher about how it was trained. Which bet would you make: the one you can verify, or the one you can't?
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 more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. I write as director of Te Pono Limited; the views are personal and do not represent the position of any client, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance, 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 "Generative AI: Skynet or Heaven" and "Space Mafia." A Concise History of Linux chronicles the operating system that changed the world and the lessons it holds for the AI era. 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: automated gates 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 are mine and are the same discipline I apply to the AI systems I audit for clients. AI accelerated the work; the thinking, and the responsibility for it, are mine.
[1] Vectara. "hallucination-leaderboard" GitHub repository, README.md. Accessed 8 October 2026. URL: https://github.com/vectara/hallucination-leaderboard/blob/main/README.md
[2] benchmarklist.com. "Vectara HHEM Hallucination Leaderboard Benchmark Scores & AI Model Leaderboard." Accessed 8 October 2026. URL: https://benchmarklist.com/benchmarks/vectara_hhem/
[3] Hugging Face. antgroup organisation page, and attempted direct access to huggingface.co/antgroup/finix_s1_32b (returned HTTP 401 Unauthorized). Accessed 8 October 2026. URL: https://huggingface.co/antgroup
[4] Hugging Face. inclusionAI organisation page. Accessed 8 October 2026. URL: https://huggingface.co/inclusionAI
[5] Paul Weiss. "EU Commission Publishes Guidelines on General Purpose AI Obligations as well as Training Data Disclosure Template." 30 July 2025. URL: https://www.paulweiss.com/insights/client-memos/eu-commission-publishes-guidelines-on-general-purpose-ai-obligations-as-well-as-training-data-disclosure-template-further-clarity-as-the-countdown-to-enforcement-begins
[6] Linux Foundation. "Open Source Summit + Embedded Linux Conference Europe 2026 Schedule Champions Open Source Innovation and Marks 35 Years of Linux." URL: https://events.linuxfoundation.org/2026/08/05/open-source-summit-embedded-linux-conference-europe-2026-schedule-champions-open-source-innovation-and-marks-35-years-of-linux/