Bijan Bowen evaluates Tencent HY3 after its full Apache 2.0 release. The mixture-of-experts modelA mixture-of-experts model contains multiple specialized subnetworks and activates a selected subset for each input instead of using every parameter every time. has 295 billion total parameters with 21 billion active, a 256K context windowA context window is the maximum amount of tokenized information an AI model can consider during one processing session. and multi-token prediction, making it a plausible open model for large unified-memory machines and multi-device hobbyist setups.
The model produces a workable first C++ skating gameCode generation uses AI or another automated system to create source code from instructions, examples, schemas, or higher-level specifications., a visually varied time-travel city, a watch storefront and a fast OpenSCAD engine shell built around a specified RC-car motor. Several results are respectable for the model's active size, but follow-up changes can make them worse, and details such as character motion, geometry and camera behavior remain inconsistent.
The most important failure comes during a Linux keyboard-lighting task. HY3 cannot identify a safe RGB control path, mistakes the display backlight for the keyboard backlight and darkens the screen before helping restore it. Bowen concludes that HY3 offers impressive cost and capability for local or inexpensive experimentationA local model runs on computing infrastructure controlled directly by the user rather than only through a remote model service., but its low-level actions still require close human supervisionHuman-in-the-loop AI keeps a person involved in reviewing, correcting, approving, or guiding an AI system's work..
Watch on YouTube
