Tencent Hy3: coding, translation and guardrail tests

This review tests Tencent Hy3 from the Hunyuan language-model family through an AI studio and a terminal workflow. The speaker describes a 295B mixture-of-experts model with 21B active parameters per forward pass, 192 experts and a 3.8B MTP layer for speculative decoding. The AI studio is hosted; the terminal test shows a local application, without establishing that the model weights run locally.

The first task asks for a self-contained HTML simulation of balls falling into an ocean, with splashes, ripples and a rainbow. The speaker judges the result slightly better than an earlier Fable test, but points out that the requested overhead perspective and 3D effect remain incomplete.

A coding assistant test then targets a full-stack call-center app whose search cannot reach the server. After the model's repair, the speaker reports a backend on localhost port 8000 and demonstrates search results and call logging. The demonstrated checks cover search and call logging.

The presenter also discusses vendor benchmark charts without independently reproducing their results. A translation test with thinking disabled produces results the speaker calls mostly correct, with some literal or inaccurate wording. The output also labels some languages as invented for illustration. A final guardrail test receives an answer with questions about the user's circumstances instead of a refusal.