Does hosting an open-weight model automatically solve data sovereignty?

No. It removes the vendor's server as a place your data has to pass through by default, which is the specific problem the sources above describe, but the deployment around it — where you host, who can access it, what logging you keep — still has to be built correctly. The licence also has to be checked: a publisher-specific text can carry its own conditions, and only a standard text like MIT or Apache 2.0 can be described from its name alone.

Are most of the models in this catalogue easy to adopt from a licensing standpoint?

Most sit under a standard permissive text with no user threshold and no field-of-use restriction, per the figure above. A smaller number carry a publisher-specific licence that has to be read before you rely on it — see /licences for the model-by-model reference, including which of each family is currently callable.

Is this saying open-weight models are more secure or higher quality than closed ones?

No. It is a claim about who controls where your data goes and what happens if a vendor changes course, not a claim about benchmark performance — see our other resources on the honest capability comparison between open and closed models.

Resources · AI Token Router

LicensingOpen-weightPricing

Data Sovereignty and the Open-Weight Advantage

Enterprises are increasingly moving away from frontier closed models specifically to keep proprietary data out of a vendor's hands, not only to cut cost [1], and data-sovereignty requirements are now cited as a reason to evaluate a deployment before any model is chosen at all [2]. Open weights make that possible in a way a closed API cannot: the weights are a file you can host wherever you choose, and they keep working even if the publisher who trained them changes terms, raises prices, or shuts down [3] — a real, separate advantage from the per-token price this site otherwise leads with.

September 8, 2026

Put your own numbers in before you take ours on trust.

Browse the full licence reference

Why sovereignty is showing up as a procurement reason, not just a cost reason

MarketScale reports enterprises actively moving away from frontier closed models to protect proprietary data, framing it as a data-protection decision rather than a cost decision [1]. NeuralTrust's enterprise guidance treats data sovereignty as something to settle before a model is chosen at all, not a checkbox filled in after — where the data physically sits, who can access it, and what leaves the organisation's control on every inference call [2].

Neither source claims open weights are categorically better models. The claim is narrower: an open-weight deployment gives an enterprise a choice about where data goes that a closed API, by its nature, does not offer at all.

What owning the weights actually changes

Forte Group states the mechanism plainly: if a provider changes pricing, deprecates a model, or shuts down, a self-hosted or independently-served open-weight deployment keeps working, because the weights themselves do not disappear when a vendor's business decisions change [3]. That is a different property from a lower price — it is a claim about what happens on the day a vendor decision goes against you, not about what you pay on an ordinary day.

A separate economic analysis on vendor switching costs, published via Zenodo in February 2026, estimated the cost of switching between major AI platforms at roughly two to six times the original implementation cost [4] — a concrete number behind the abstract idea that vendor lock-in is expensive to escape. Owning the weights is one way to avoid ever needing to pay that switching cost in the first place.

Rate-card discount, callable models

Owning the weights is a separate argument from price, and the two happen to point the same way here -- the same catalogue that keeps working on its own terms also prices below what each model's own publisher charges. Kimi K2.6 carries the deepest discount in this set at 42%, against 36% for DeepSeek V4 Flash. Figures are the rate-card discount, with no caching assumed.Computed from our published rates against the official rate.

Where our catalogue's licences actually sit

None of that is automatic just because a model is described as "open weight." Licence terms vary by publisher, and only a standard, unmodified text — MIT, Apache 2.0 — can be described from its name alone; a publisher's own licence document has to be read before you rely on it. Most of the models in this catalogue carry a standard permissive text with no user threshold and no field-of-use restriction. A smaller number carry a publisher-specific licence whose terms are set by that document rather than by anything inferable from its name — the segment below is computed once from the licence recorded against every model, not asserted.

Licence families across the catalogue

24 models

Apache 2.01146%
MIT729%
Modified MIT28%
Llama 4 Community14%
LTXV Open Weights14%
Stability Community14%
Tencent Hunyuan Community14%

Apache 2.0 and MIT together cover most of the 24 models in the catalogue and require nothing of a commercial user beyond keeping the licence notices; the rest sit under five publisher-specific texts, each of which has to be read on its own before you rely on it for a sovereignty-driven deployment.Computed once from the licence recorded against each model in catalog.ts — see /licences for the full, always-current reference.

What this does not claim, and what to check before you act on it

This is not a claim that a sovereignty-driven deployment is free of work. A publisher-specific licence still has to be read, not assumed, and not every model under a given licence family is currently callable through this platform — check the model's own page for whether it is servable today before you plan a deployment around it. And sovereignty is a real, cited reason enterprises choose open weights [1][2]; it is not evidence that open-weight models match or exceed closed frontier models on capability, a separate question this site does not settle here.

Sources

[1] Enterprises are ditching frontier AI models for open-source alternatives to protect proprietary data — MarketScale. Cited for the claim that data protection, not price, is driving some enterprises away from frontier closed models.
[2] AI Data Sovereignty for Enterprise: Why It Matters Before You Deploy LLMs — NeuralTrust. Cited for treating data sovereignty as a pre-deployment decision rather than an afterthought.
[3] Owning the Stack: What Open Weight Models Change for Enterprise — Forte Group. Cited for the mechanism: a self-hosted or independently-served open-weight deployment keeps working after a vendor's own pricing, deprecation or shutdown decision.
[4] Zenodo record: switching-cost analysis across major AI platforms (Feb 2026) — Zenodo. Cited for the estimated 2.3x-5.7x switching cost between major AI platforms, as context for why avoiding lock-in has a real, estimated price.

Questions this raises

Does hosting an open-weight model automatically solve data sovereignty?: No. It removes the vendor's server as a place your data has to pass through by default, which is the specific problem the sources above describe, but the deployment around it — where you host, who can access it, what logging you keep — still has to be built correctly. The licence also has to be checked: a publisher-specific text can carry its own conditions, and only a standard text like MIT or Apache 2.0 can be described from its name alone.
Are most of the models in this catalogue easy to adopt from a licensing standpoint?: Most sit under a standard permissive text with no user threshold and no field-of-use restriction, per the figure above. A smaller number carry a publisher-specific licence that has to be read before you rely on it — see /licences for the model-by-model reference, including which of each family is currently callable.
Is this saying open-weight models are more secure or higher quality than closed ones?: No. It is a claim about who controls where your data goes and what happens if a vendor changes course, not a claim about benchmark performance — see our other resources on the honest capability comparison between open and closed models.

AI Token Router is an OpenAI-compatible gateway for open-weight models, priced below each publisher’s own rate on every row.

Browse the full licence reference Get $5 in free credits

What a Fixed Monthly AI Budget Actually Buys in 2026
Three realistic budget tiers, worked by hand against this catalogue's own rate table, at one stated request shape -- how many requests and tokens $10, $50 and $250 a month actually buys on three callable models.
When a Closed Frontier Model Is Still the Right Call
Closed frontier models measurably lead reasoning-heavy benchmarks as of September 2026. Where that lead and a simpler operational model are worth the higher price -- and why our catalogue is not the answer for that reader.
Self-Hosting vs a Managed Open-Weight API: When Each Wins
Where the self-host breakeven actually sits, what self-hosting really costs once engineering time is priced in, and the honest cases where self-hosting wins -- this is not a blanket argument for a managed API.