Descriptions:
Fahd Mirza puts Tencent’s newly released Hunyuan 3 (Hi3) through a battery of practical tests in Tencent’s AI Studio, covering physics simulation generation, full-stack application debugging, and multilingual translation. Hi3 is a 295-billion-parameter mixture-of-experts model with 21 billion active parameters per forward pass across 192 experts, plus a 3.8-billion-parameter MTP layer for speculative decoding — a significant upgrade over the April preview release, with post-training scaled on higher-quality data drawn from feedback across 50 internal products.
On benchmarks, Hi3 clears meaningful gains over its own preview and consistently beats GLM 5.2 and DeepSeek V4 Pro across most evaluations. Its strongest results appear on SWE-Bench Pro, Terminal Bench 2.1, and BrowseComp; its weakest relative showing is on Math Arena Apex, where it trails GLM 5.2, Seed 2.1 Pro, Qwen 3.7 Max, and DeepSeek. Claude Opus 4.8 and GPT-5.5 remain ahead overall, though Mirza notes the gap is smaller than expected.
In live testing, Hi3 successfully diagnosed and fixed a multi-bug full-stack call center application from raw code with no additional context, produced a physics simulation that slightly edged Claude Fable 5 on the same prompt, and handled multilingual translation across dozens of languages with generally solid accuracy. Mirza also briefly probes the model’s guardrails before wrapping up.
📺 Source: Fahd Mirza · Published July 06, 2026
🏷️ Format: Review
