Bijan Bowen reviews Tencent HY4 Preview as a 770-billion-parameter 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. with 49 billion active parametersAn active AI model parameter is a model weight selected to participate in computation for a particular token or inference step., a one-million-token context window, Apache 2.0 weights and a built-in speculative-decoding layer. Tencent positions it for long-horizon software engineering and document-heavy work, while acknowledging that the preview can overthink and oververify tasks.
The hands-on tests cover a browser operating system, a skateboard game, a subway shooter, an interactive watch page, a Blender city timeline and a retro wrestling game. HY4 often produces rich details, functional interactions and ambitious visual scenes, but several outputs need targeted fixes, omit requested behavior or show obvious preview-stage defects.
Bowen's strongest criticism is the serving experience. Frequent 429 rate-limit errorsAPI rate limiting restricts how many requests a client can make within a defined period or resource budget. interrupt long agentic runsAI agent reliability is the degree to which an agent consistently completes intended work correctly and safely under expected conditions. even after he supplies billing details, making the hosted model unreliable for sustained workModel reliability is the degree to which an AI model produces dependable behavior across repeated, varied and operationally realistic use.. He remains optimistic about the full release and a possible smaller variant, but treats the current preview as promising rather than production-ready. Subscription and comment requests are omitted.
Watch on YouTube
