A Heidelberg laboratory handed the world an open-weight model on 3 October 2026 and called it sovereign. The weights are free to download under Apache 2.0, trained in Germany and Finland, and runnable on hardware the buyer controls. Two weeks earlier, the same company signed an agreement to combine with a Canadian firm. The release is real and the licence is genuinely permissive. The ownership of the entity that maintains the model is not settled, and that is the part enterprise buyers have not priced. Sovereignty was given away as a download while control of the vendor moved across an ocean.
What the original report says
On 3 October 2026, Aleph Alpha announced Kolibri, an English-German Mixture-of-Experts language model with 78.1 billion total parameters and 3.46 billion active per token, released under the Apache 2.0 licence with full weights on Hugging Face. The Heidelberg company said the model was built in Germany and trained on infrastructure in Germany and Finland “with no foreign control,” aimed at regulated work in public administration, industry and aerospace. The 189-page technical report describes a mixture-of-experts transformer with a native context of 262,144 tokens and a recommended ceiling at that length for complex tasks. Aleph Alpha has signed the EU General-Purpose AI Code of Practice and published a training content summary. On 16 September 2026, Aleph Alpha and Canada’s Cohere signed a definitive business combination agreement under which the combined company would operate as Cohere, with dual headquarters in Toronto and Berlin, still subject to regulatory approval.
Featured image: Aleph Alpha’s Kolibri release announcement, published 3 October 2026. Credit: Aleph Alpha.
THE SCENARIO: A ministry in a mid-sized country has just cleared a two-year programme to run a public benefits model in-house. The procurement team chose an open-weight model precisely because it can be downloaded and run on national hardware. The business case said the state would never again depend on a foreign vendor’s roadmap. Then the vendor signs a combination agreement with a company in another jurisdiction. The model files do not move. The maintenance relationship, the security fixes, the next release and the terms of service all do. Nobody in the ministry can name the entity that will answer the support ticket in eighteen months.
The Question
The question is not whether open weights are a genuine sovereignty tool. They are, and the release on 3 October is a more serious contribution than most sovereign AI announcements deliver. The question is what a state or an enterprise actually holds when it downloads a model it cannot maintain, evaluate or replace on its own terms, from a vendor whose ownership is changing.
A month of announcements makes the tension visible. On 3 October Aleph Alpha published the full Kolibri weights for free. On 16 September the same company signed a definitive agreement to combine with Cohere, with the combined entity operating as Cohere and headquartered in Toronto and Berlin. On 29 September a national telecoms operator in that vendor’s market signed a memorandum of understanding with a network vendor to deliver sovereign infrastructure, an arrangement whose management plane has not been made public. Every release in the sequence uses the word sovereign. None of them answers the question a board actually has to answer.
For leaders, the practical issue is which passport is on the contract, and which government can compel the entity that holds the keys. That matters because sovereignty is not a characteristic of a file. It is a property of a relationship that survives a change of control, a licence change, a support decision and a regulator’s demand. A model on a disk is an asset. A model on a disk with no one contractually obliged to fix it is inventory with a shelf life. The question this article examines is whether a free download can carry a state’s sovereignty ambition when the vendor behind it is in the middle of a merger.
What Happened and Why It Matters
On 3 October 2026, timed to the Day of German Reunification, Aleph Alpha released Kolibri, an English-German mixture-of-experts language model, with the full weights published on Hugging Face under the Apache 2.0 licence. The company’s own announcement describes a model of 78 billion total parameters with roughly 3 billion active per token, a context window running to a million tokens, and a design target of sovereign mission-critical work in public administration, industrials and aerospace. The model was built in Germany and trained on infrastructure in Germany and Finland. Aleph Alpha says it owns the entire pipeline, from data curation through pre-training and post-training to final evaluation, and that a customer can run the model on hardware it controls rather than sending data to an outside inference service.
That is a substantive release, and it deserves to be described as one. The licence is the standard Apache 2.0 text, verified against the Apache Software Foundation’s copy by the independent review site SignalStack. The weights are not gated. The model can be downloaded, fine-tuned and shipped, and it needs roughly 78 gigabytes of memory for its FP8 weights plus a vLLM plugin supplied by the vendor. For a European public body whose binding constraint is a regulator rather than a leaderboard, the compliance framing is the load-bearing part: the release targets the EU AI Act, the General-Purpose AI Code of Practice and the GDPR, and the vendor published a training content summary in the AI Act template.
The scrutiny that followed is equally important, because it shows where the sovereignty claim is strong and where it is thin. Independent checks by The Frontier and Traictory found that the Apache 2.0 grant covers only the weights and configuration files in the repository, not the training code, the architecture, the parameter settings or the training methods. Buyers can run and adapt the model. They cannot inspect how it was built beyond what the vendor chose to publish. The benchmark table in the release is also the vendor’s own run: Aleph Alpha scored its model against Qwen, Nemotron and Mistral rivals, and the headline claim of matching models up to four times its active parameter count rests on a narrow set of reasoning and knowledge rows rather than the aggregate. On the vendor’s own hallucination index, Kolibri scores worse than the smaller Qwen rival in the same table. None of this makes the release dishonest. It means the evidence for the claim is produced by the party making it.
The part that matters most for institutional buyers is a corporate fact, not a technical one. On 16 September 2026, Aleph Alpha and Cohere signed a definitive business combination agreement. The combined company is to operate globally as Cohere, with dual headquarters in Toronto and Berlin, and the transaction remains subject to regulatory approval in both countries. Until closing, Aleph Alpha continues to operate independently. The Kolibri release was prepared and published inside that window. The model carries the phrase sovereign AI made in Germany. The company that made it has agreed to become part of a firm headquartered partly outside Europe.
That matters because the sovereign artifact and the sovereign relationship are not the same thing, and only one of them was given away. Analysis published by theCUBE Research on 7 October put the sequence together and reached the conclusion that a buyer can take home exactly one thing from the month’s announcements: the weights. The platform, the alliance, the infrastructure memorandum and the support relationship are all delivered on terms that a statement of work can change. The same analysis noted that a national compute build in Canada was operated by a United States company, so that the soil was domestic while the landlord filed with a foreign regulator. The pattern is not a scandal in any single case. It is a structural feature of the market, and it is the feature a sovereignty strategy has to budget for.
The distinction generalises. In an open-weight release, the buyer receives a finished object. The vendor retains the upbringing: the data lineage beyond the published summary, the evaluation harness, the security review process, the roadmap and the decision rights over the next version. The technical report accompanying Kolibri runs to 189 pages and is unusually candid for the genre. It still does not transfer the capability to maintain the model forward. When a state adopts such a model for a public function, it inherits operational responsibility for an artifact whose future is decided elsewhere, unless it separately builds the capability to inspect, evaluate and replace it. The release makes that possible. It does not make it automatic.
Two other developments in the same four-week window show that the pattern is being recognised across jurisdictions. On 29 September, a national telecoms operator announced a memorandum of understanding to deliver sovereign AI infrastructure for domestic customers, built on a foreign vendor’s stack that also holds a stake in the model company. A memorandum of understanding is a promise to keep talking; it does not settle where the management plane sits, who holds the keys, or who gets paged at three in the morning. That matters because a sovereignty claim that lives only in a press release is a claim nobody has tested. The practical test arrives at the first renewal, the first change of model version, and the first regulator’s request for the audit trail.
The evidence from this month also shows what a credible version of the claim looks like. The Kolibri release passes cleanly on the territorial question, because the weights land wherever the buyer puts them. It passes on auditability for the weights, because they can be self-hosted and forked. It is conditional on the operational question, because the pager belongs to whoever signed the support contract. It is incomplete on the legal question, because a company with two headquarters answers to two sets of statutes and two regulators, and the combination still needs approval from both governments. And it is unpriced on the financial question, because the platform economics of a firm reported to be valued around twenty billion dollars will be repriced when it seeks public markets. Four of five categories examined well. One category decided the outcome.
For leaders, the practical issue is therefore not whether to adopt open weights. The practical issue is what else has to be in place before an open-weight adoption counts as sovereignty rather than a cheaper form of dependency. A downloaded model with no local evaluation capability, no maintenance plan and no tested exit is a rented capability with a permanent-looking file on disk. A downloaded model with a standing evaluation function, a documented fallback and a contract that survives a change of control is a strategic asset. The gap between the two is entirely within the control of the buying institution, and it is the gap this article turns to next.
The Sovereignty Risk
The risk in an open-weight adoption is a transfer of apparent control without a transfer of actual control. The buyer receives the artifact and the visible part of the capability, and loses the invisible part: the ability to maintain, evaluate, correct and replace the system on its own timeline. That matters because the invisible part is where operational risk accumulates. A model that cannot be independently evaluated in the buyer’s own language and use cases cannot be governed. It can only be trusted or distrusted, and neither is a governance position.
The first dimension of the risk is dependency that survives a permissive licence. A model that is open-weight but tightly coupled to a vendor’s inference plugin, serving stack and evaluation harness is not removable without dismantling the applications built around it. Independent reviews of the Kolibri release made the coupling concrete: the model requires the vendor’s own package to serve, and the licence grant stops at the weights. The file is free. The operating relationship is not, and the operating relationship is where the switching cost lives. The sovereignty test is whether the organisation can replace the model without rebuilding the services, the security policies and the retrieval systems around it.
The second dimension is the accountability gap created by a change of control. When a vendor signs a combination agreement, the buyer’s contract, support path, data processing terms and roadmap all pass to a new signing entity. A company with dual headquarters in two jurisdictions answers to two regulators and may be subject to legal process in either. Analysis published by theCUBE Research framed the consequence precisely: two passports are not two shields, they are two subpoenas. For a public body that adopted a model on a compliance argument, a change in the identity of the counterparty is a change in the compliance position, and it can arrive without any change to the technology the buyer uses.
The third dimension is the loss of institutional capability that follows from outsourcing maintenance. A state that buys a model and rents the expertise around it has not built sovereign capability; it has bought a longer lease. The skills to evaluate a model against national benchmarks, in national languages, on national use cases, do not arrive with the download. When the integrator leaves, the evaluation function, the agent memory, the audit logs and the incident response all leave with it unless the contract placed them in the buyer’s tenancy. That matters because an organisation may work with partners to deliver those capabilities, but it cannot outsource accountability for them.
The fourth dimension is the price risk that open weights do not remove. The model is free. The platform around it is priced by a commercial firm whose investors expect a return, and a firm reported to be valued near twenty billion dollars and considering public markets has a forcing function that will eventually reach its customers. Modelling unit economics at today’s platform price without a stress case is a governance failure, not an optimism. The question at renewal is whether the buyer can leave on its own timeline, at a survivable cost, without a meter it does not own setting its margin three years out.
The TEE Method Response
The TEE Method addresses sovereignty through three assets that an institution must hold at home: Talent, the human capability to evaluate and operate; Enterprise, the contractual and operational control that determines who decides; and Ecosystem, the local market and supply relationships that give the institution alternatives. Applied to an open-weight adoption in a changing corporate landscape, the method converts a technology decision into an institutional one.
On Talent, the release makes one thing easier and one thing harder. It is easier to build capability because a capable bilingual model can be downloaded, run air-gapped and fine-tuned without negotiating vendor access. It is harder because the skill that matters is no longer running the model but judging it. The capability to hold at home is the ability to independently measure model behaviour against the institution’s own benchmarks, in its own languages and use cases, rather than accepting the vendor’s scorecard. Aleph Alpha’s own benchmark table is the demonstration: every figure in it was produced by the party with an interest in the result. A public body that cannot reproduce those numbers on its own tasks has not evaluated the model. It has read a press release with a chart.
On Enterprise, the practical work is contractual and administrative. Every AI contract renewing in the next twelve months should carry a change-of-control clause that requires the buyer’s consent for assignment and holds pricing for a defined period after any combination closes. The signing entity should be named in the risk register next to the statutes it brings. The crown jewels, meaning identity, audit logs, agent memory and context, and the evaluation harness, should be placed in the buyer’s tenancy in the statement of work and transfer to the buyer’s team on the day the integrator leaves. These are unglamorous clauses. They are the mechanism by which a sovereignty claim survives contact with a corporate transaction.
On Ecosystem, the release is genuinely useful because it widens the set of suppliers outside the two dominant frontier markets. A third option with permissive weights and a European compliance posture is better for every regulated buyer than one fewer alternative. But a vendor can sell sovereign components; it cannot sell sovereignty. The ecosystem work is to convert the availability of an alternative into a functioning domestic market: local integrators who can maintain the stack, national evaluation capacity that can test it, and enough domestic demand to sustain both. That is what turns a free download into a durable position, and it is the work that a procurement announcement does not do by itself.
Sovereignty Test Matrix
The matrix scores five sovereignty domains from one to five, using the signals that appeared in this month’s announcements. A weak score means the institution depends on another party’s decision. A stronger score means the institution retains the ability to inspect, test, correct, export, interrupt and replace what it runs. The point of scoring is to make the trade-offs visible before the money is committed, not after.
| Domain | Weak signal | Stronger signal | Score |
|---|---|---|---|
| Political | Model adopted on a vendor’s compliance claim with no national evaluation body | National capability to assess the model against domestic law and policy | 3 / 5 |
| Economic | Free weights, priced platform, no stress case on post-transaction pricing | Costed exit and fallback benchmarked at a multiple of today’s platform price | 2 / 5 |
| Cultural | Model’s quality in national languages asserted by the vendor | Independent testing in the buyer’s own languages and use cases | 3 / 5 |
| Intellectual | Weights open, training pipeline, evaluation harness and security process closed | In-house evaluation and the ability to fork and correct the model | 3 / 5 |
| Technological | Vendor plugin required to serve, no tested replacement path | Air-gapped deployment on national hardware with a tested exit before adoption | 4 / 5 |
The pattern in the scores is the article’s argument in one line. The territorial and technological domains are strong, because the weights can be held and run locally. The economic and intellectual domains are weaker, because the maintenance, evaluation and pricing decisions sit with the vendor and its future owner. A programme that scores well on technological control and poorly on economic exit is a strong technical position with an unhedged commercial exposure. A programme that scores well on economic exit and poorly on cultural fit can be secure and still wrong, because a model that does not work in the languages and services it must serve is not serving the population.
Two scores deserve elaboration. The cultural score of three reflects a real achievement understated by its own vendor. Aleph Alpha trained the model with German and English as first-class languages, reporting that German accounted for a material share of pre-training tokens, and built abstention behaviour into the model so it declines to answer when context does not support a claim. That is a genuine contribution to linguistic sovereignty, and it is the opposite of the pattern in which a model is trained in one language and translated into others. The intellectual score of three reflects the limit of the licence: independent reviewers confirmed the grant covers the weights and configuration files but excludes the training code, architecture and methods. The buyer gets the finished object and not the upbringing.
What Leaders Should Do This Week
The actions below are specific enough to assign to a named owner and complete inside a planning cycle. None requires new technology. All of them address the gap between holding an artifact and holding control.
First, download the weights and run your own evals on your own tasks. Even if the model is never deployed, a benchmark the institution owns changes the negotiating position at the next renewal. This is assignable to the evaluation or data science lead and can be completed in days, because the weights are ungated.
Second, add a change-of-control clause to every AI contract renewing in the next twelve months. The clause should require the buyer’s consent for assignment and hold pricing for a defined period after any combination closes. The legal and procurement leads should own this, and the clause should be written before the next renewal rather than after.
Third, name the signing entity in the risk register and record the statutes it brings. Toronto, Berlin or another jurisdiction is not a formality; it determines which regulator and which legal process reach the counterparty. The contract owner should document this for every critical AI supplier.
Fourth, establish where the management plane sits. Ask the infrastructure provider who holds the keys, who gets paged, and who can pull the plug. Where the answer is that it depends, ask what it depends on and whom it depends on. An infrastructure or security lead should obtain this in writing.
Fifth, put the crown jewels in the statement of work. Identity, audit logs, agent memory and context, and the evaluation harness should sit in the institution’s tenancy and transfer to its own team on the day any integrator rolls off. The procurement lead should make this a condition of award rather than a preference.
Sixth, stress case the platform price. Model unit economics at three times today’s platform price and test whether the margin survives it. A private-company price is not a durable price, and the finance lead should treat the multiple as an assumption to be documented rather than a risk to be noted.
Seventh, test the exit before adoption, not after. Confirm that the model can be replaced without rebuilding the applications, security policies and retrieval systems built around it. The test should be run in a staging environment and its result recorded as a decision threshold: if replacement requires a rebuild, the architecture has created lock-in whatever the licence says, and the programme should not proceed on a sovereignty argument until that is remedied.
The Question Revisited
The question was whether a free download can carry a state’s sovereignty ambition when the vendor behind it is in the middle of a merger. The answer from the evidence of the past month is precise. It can carry the artifact. It cannot carry the relationship. Aleph Alpha’s Kolibri release is a real contribution: permissive weights, European training, a compliance posture built for regulated buyers, and a serious commitment to a second language rather than a translated afterthought. Those are the things a buyer can take home and keep, and they are more than most sovereign AI announcements deliver.
Read through the TEE Method, the strength and the exposure fall into place. Talent is the domain the release makes possible and the buyer must still build: the ability to evaluate the model independently in the institution’s own languages and use cases. Enterprise is the domain where the gap sits, because the contract, the support relationship and the change of control determine who decides after the download. Ecosystem is the domain that turns an available alternative into a functioning market, and it depends on local integrators, national evaluation capacity and domestic demand that no press release creates.
That matters because the sovereignty test is not whether the model is open. It is whether the state can inspect, test, correct, export, interrupt and replace what it runs, and whether that ability survives the vendor’s next transaction. A ministry that downloads the weights and stops there has achieved a degree of control and not sovereignty. A ministry that downloads the weights, builds the evaluation function, writes the change-of-control clause and tests the exit has converted a free artifact into a durable position. The difference between those two outcomes is a page of contract terms and a benchmark the institution owns. The question is not whether to adopt open weights. It is what has to be true for the adoption to still be yours in eighteen months.
For boards and public leaders, there is a useful framing to hold. A vendor can sell sovereign components. It cannot sell sovereignty. The part a buyer actually owns is the part nobody else can revoke: the data, the evaluation, the exit and the right to answer to its own regulator. Those are the assets that transfer when the underlying model changes, and they are the only ones worth budgeting for as if they matter, because they do.
Editorial note: This is a news-led Sovereignty Briefing analysis produced with Hermes-assisted research and source checks.
Sources
“Kolibri Has Landed: A Sovereign Open-Weight Model” — Aleph Alpha, 3 October 2026.
“Aleph Alpha releases Kolibri — Sovereign AI made in Germany” — Aleph Alpha, 5 October 2026.
“Sovereign AI Just Got Dual Citizenship” — Amit Eyal Govrin, theCUBE Research, 7 October 2026.
“Aleph Alpha Kolibri: License, Specs, Benchmarks” — SignalStack, 4 October 2026.
“Aleph Alpha Kolibri-1 opens 78B German and English MoE weights under Apache 2.0” — The Frontier, 4 October 2026.
“Aleph Alpha’s Kolibri open weights: a 78B German MoE and a vendor’s own scorecard” — Traictory, 5 October 2026.
“Aleph Alpha Releases Kolibri as the Cohere Deal Awaits Approval” — Grace Sharp, MRKT3.0, 5 October 2026.
This article draws on the TEE Method™ framework from SOVEREIGN: Who Owns the Future? For the complete framework and the full sovereignty diagnostic, see tonishatagoe.com.
Sign in to save