The licence is the reason this is possible, so we publish it

The question legal actually asks

The blocking question in a model procurement review is rarely about price and almost never about benchmarks. It is: under what right is this vendor serving these weights to us, and what happens to our data once it reaches them. Both have documentary answers, and a vendor who cannot produce them quickly is a vendor whose review stalls for a quarter.

So the answers are published rather than available on request. Every model in the catalogue names its licence on its own page. The models we refuse to serve are listed too, with the reason for each refusal, because a list of what a vendor will not do is more informative to a reviewer than a paragraph asserting that the vendor is careful.

The commercial reason for this alignment is worth stating plainly: an open licence is what makes third-party hosting lawful, third-party hosting is what puts the cost base under our control, and control of the cost base is what produces the price difference. A vendor reselling closed-weight access at a discount would have neither the right nor the margin, which is why we do not do it.

The catalogue, with licences attached

Below is every model we can serve, with its publisher, its licence and its current availability. The licence column is the one a reviewer reads first; the availability column is the one an architect reads first.

Pricing is published on the same table as the licence rather than in a separate quote, and every rate sits 36–42% below what the model's own publisher charges. Nothing on this page is contingent on a procurement conversation.

Availability is stated in the table rather than discovered at call time. A model marked unavailable has a published price and no upstream serving it yet; requesting one returns an error naming the model rather than a substitution.

The cost side of the business case

A licence review usually runs in parallel with a cost approval, and the second is easier to produce evidence for. The calculator prices your volume across every model we serve against the publisher's own rate.

The result is a shareable URL, which is the format a finance approver can check rather than take on trust. It is also the format that survives being pasted into a procurement ticket.

What a review needs from us, in order

Three documents, all of them already published. None of them requires a call, an NDA or a sales conversation to obtain.

1. 1

   The licence for each model you intend to use

   Named on the model's own page and in the licence reference, with the family explained. A reviewer needs the licence name and the permission it grants for third-party hosting and for commercial use of the outputs.

2. 2

   The data-handling schedule

   What is stored, for how long, and what is never stored. Chat and completion traffic is relayed and not retained. Asynchronous job inputs and results are held for 30 days so they can be fetched, or deleted immediately on request. Usage metadata is retained because the invoice is built from it.

3. 3

   The refusal list

   The models we do not serve and why. A reviewer assessing supply-chain risk learns more from a specific refusal than from a general assurance, and this is the list that shows the policy is applied rather than stated.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://router.xark.io/api/v1"   # ← the only line that changes
)

response = client.chat.completions.create(
    model="z-ai/glm-5.2",
    messages=[{"role": "user", "content": "Hello"}]
)

Licence and privacy questions that are not answered on the published pages go to affiliate@xark.io, which is a monitored address rather than a role alias that bounces.

The models we refuse, and why

This list exists because it is the most useful thing a vendor can show a reviewer. Each of these is a model customers ask for, and each is one we will not serve.

Sora 2 — OpenAI

Closed weights — no resale license exists

Veo 3.1 — Google DeepMind

Closed weights — no resale license exists

Runway Gen-4 — Runway

Closed weights — no resale license exists

Kling 2.5 — Kuaishou

Closed weights — no resale license exists

Wan 2.5 / 2.6 — Alibaba

API-only — weights never released, despite the open Wan 2.1/2.2 lineage

What we do with your data

We do not train on prompts or completions, ever. Chat and completion traffic is relayed to the model and not stored by us at all. The inputs and results of asynchronous jobs — video and image generation, which cannot return inline — are held for 30 days from completion so you can fetch them, or deleted immediately on request.

Usage metadata, meaning the model, token counts, latency and cost, is retained for 24 months because the dashboard and your invoice are built from it. The privacy policy carries the authoritative schedule, and where this page and that page differ, that page is the one to cite in a review.

Contracting entity

Innoprise Incubator INC, named identically in the terms, the privacy policy and the compliance page so a reviewer is not reconciling three documents.

Output rights

Determined by the licence of the model that generated them, which is why the licence is on the model page rather than abstracted into a single site-wide statement. Licence families differ on this and the difference matters.

No silent substitution

A request for a specific model is served by that model or fails. Substituting a model would change the licence terms of the output you receive, which is a compliance event and not merely an engineering one.

Procurement and licensing questions

The questions a review actually blocks on, answered without hedging.