Side-by-side comparison of Hy3-preview (Tencent) and Mixtral 8x22B Instruct (Mistral AI · France) for self-hosted deployment of the open-weight model. Hy3-preview is rated blocked; Mixtral 8x22B Instruct is EU-ready. They part ways on licence: Hy3-preview is "Tencent Hy Community", Mixtral 8x22B Instruct is "Apache 2.0".
| Field | ||
|---|---|---|
| Summary | ||
| Verdict | Blocked Per the Tencent Hy Community License Agreement (Sections 1(l) and 5(c)), the licence's defined Territory excludes the European Union, the United Kingdom and South Korea, and licensees are expressly prohibited from using, distributing or displaying the model or its outputs outside that Territory. Under those terms the weights are not deployable for EU users or workloads, regardless of architecture quality. | EU-ready A rare combination of permissive licensing, EU provenance, and credible multilingual coverage — Apache 2.0 leaves commercial deployment essentially unconstrained, and Paris-based Mistral simplifies the GDPR controller story compared with US-hosted peers. The catch is the training corpus: Mistral has never published a meaningful dataset breakdown, so AI Act transparency obligations on training-data summaries land squarely on the deployer. |
| Last reviewed | 2026-04-28 | 2026-05-19 |
| Open-weight | ||
| Licence | Tencent Hy Community | Apache 2.0 |
| Commercial use | EU territory excluded | Unrestricted |
| Training data | Undisclosed | Undisclosed |
| Origin | China | EU (France) |
| Performance & pricing? | ||
| Quality index | 42/100 | — |
| Speed | 85 tok/s | — |
| Blended price | — | — |
| Context window | — | — |
| Evidence | ||
| Sources | ||
No overlapping sources between the two entries.