In a move that reshapes the availability landscape for open models, Tencent has released the Hy3 LLM under an Apache 2.0 license, making it usable even in regions that were explicitly excluded until yesterday. The collection is available on HuggingFace and marks a clear departure from the previous community license, which forbade use in South Korea, the United Kingdom, and the European Union.

Hy3 is a massive model: 295 billion total parameters, yet only 21 billion active for each generated token — a figure that reveals a mixture-of-experts (MoE) architecture. This design, familiar to those tracking large open models, keeps inference costs manageable while retaining a broad knowledge base within the total parameters. On the hardware side, the 21 billion active parameters suggest that inference can be handled with high-end GPUs without requiring sprawling clusters, although the full model demands enough VRAM to host all experts, making it a non-trivial load for fully on-premise deployments. Techniques like quantization can shrink the footprint, but the trade-off between quality and hardware requirements remains central.

The real turning point is the license. The shift to Apache 2.0 erases the geographical restrictions that kept the European and British markets out. For organizations subject to GDPR and with data sovereignty needs, this means being able to evaluate Hy3 in self-hosted or air-gapped scenarios without risking contractual violations. The license opening also places the model within a broader trend: even large Chinese vendors are starting to release weights under permissive licenses, reducing reliance on external cloud providers and widening the options for those building local inference stacks.

Anyone planning an on-premise deployment knows that choosing a model is just one piece of a puzzle that includes hardware, serving frameworks, and operational costs. The availability of an LLM with a relatively small active footprint and a license without geographical constraints changes the contours of the problem: engineering challenges remain, but the legal uncertainty that served as a preventive stop for many European companies is gone. AI-RADAR explores the trade-offs of open models in on-premise contexts through dedicated analyses, precisely to support decisions in scenarios like this one.