Loading the catalogue…
Loading the catalogue…
US/Korean inference provider serving open-weight models via serverless Model APIs over an OpenAI-compatible API.
Five dimensions
Bands only. The scores and Stav's notes per dimension are read in the workspace.
Stav's narrative, every risk item with its source, the safeguards that offset them and the privacy-policy issues found: the part of the page a procurement review runs on.
Stav's narrative · 5 risk items · 3 safeguards · DPA, sub-processor list, terms, privacy policy — sign in to read them.
Traffic through Stav is covered by Stav's customer DPA, which lists Stav's sub-processors.
This level applies to traffic routed through Stav.
FriendliAI is at L0. L1 is blocked because data and infrastructure are in the US on AWS, Oracle and Google, and FriendliAI Corp. can be compelled under the CLOUD Act and FISA.
3 criteria from L1: Customer data in the Union, Infrastructure in the Union, No coerced vulnerability silence.
Stav's own assessment against criterion registry 4.1. Indicative, not an official recognition, which only a national authority can confer. Criteria follow the proposed Cloud and AI Development Act and are expected to change. “Union” reads as EU/EEA at every rung. CADA itself still says “in the Union”: until the Act is incorporated into the EEA Agreement, a Norwegian, Icelandic or Liechtenstein route does not evidence formal Union assurance for an EU public-sector buyer. UK and Swiss providers are capped at L1.
Verified 17 Oct 2026About the ladder →“Claimed by the provider” means the provider states the certification. Stav has not seen the certificate or report.
| Model | Input / M | Output / M | TTFT | Tok/s |
|---|---|---|---|---|
| GLM-5.3 Flash | €0.13 | €0.45 | 1.46 s | 140 |
| GLM-5.3 | €1.13 | €3.55 | 0.48 s | 378 |
| glm-5-2 | €1.26 | €3.95 | 0.45 s | 304 |
| GLM-5.1 | €1.26 | €3.95 | 0.52 s | 163 |