Summary
Fahd Mirza puts Tencent’s HunyuanLarge3 (Hi3) Preview through a series of demanding real-world tests shortly after its April 23, 2026 release. Hi3 is a 295-billion-parameter mixture-of-experts model that activates only 21 billion parameters at inference time, supports a 256K context window, and includes an adjustable built-in reasoning mode. Weights are publicly available on Hugging Face and ModelScope, and the model is currently accessible for free via OpenRouter without an API key.
Mirza runs three escalating challenges: generating a self-contained HTML/WebGL reaction-diffusion simulation from a single prompt — a task he frames as equivalent to architecting a real-time physics rendering engine — then acting as a research agent with live web search to produce a supply chain narrative tracing a lithium battery cell from cobalt mines in the Democratic Republic of Congo through Chinese chemical processing to consumer devices. The supply chain output earns particular attention for including specific dollar figures, chemical compound names, legislation references, and named individuals woven into narrative prose rather than bullet points. A third test examines the model’s ability to reason its way out of a logically constrained scenario.
Benchmark slides shown in the video place Hi3 Preview competitively against significantly larger models on International Math Olympiad problems, frontier science tasks, and real-world software engineering evaluations. For developers and researchers tracking the open-source model landscape, this is an early hands-on look at a serious new entrant from Tencent’s rebuilt training infrastructure.
📺 Source: Fahd Mirza · Published April 26, 2026
🏷️ Format: Review
