Tencent Hy3 — technical and market analysis

Technical analysis · July 2026 Tencent Hy3: architecture, benchmarks and significance for the open-source model market Technical and market analysis of the Mixture-of-Experts model from Tencent AI Lab Published: July 2026 · Sources: Hugging Face, Artificial Analysis, Yahoo Tech / Decrypt Introduction On July 6, 2026, Tencent released Hy3 as an open-source model — the full version of the successor to Hy3 preview, which debuted in preview mode three months earlier, on April 23.

The release went relatively unnoticed in the Western tech environment, but the numbers it contains are hard to ignore: a model with 295 billion total parameters and only 21 billion active during inference achieves results comparable to models three times larger — at a fraction of their computational cost.

Hy3 is a product of the Tencent Hy Team (formerly Hunyuan), the research group responsible for the Hunyuan series within one of the largest Chinese technology corporations.

Releasing the model under the Apache 2.0 license — allowing for free commercial use — signals Tencent's ambitions beyond the Chinese market and a direct strike at the open-source ecosystem dominated until now by Meta (Llama), Mistral, and DeepSeek.

Context: why Hy3 is important To properly evaluate Hy3, it is worth looking back one year.

When DeepSeek released the R1 model in 2025, the market took it as a sign that the era of Western laboratories' monopoly on frontier models was slowly coming to an end.

DeepSeek showed that a frontier-class model can be built with a significantly lower computational budget than previously assumed.

Hy3 takes it one step further in this same logic: not just cost efficiency, but parametric efficiency.