Tencent open-sources the Hy4-preview mixture-of-experts model

Large language models and developer tools | China/global | 28 August 2026 — 72-hour extension

Tencent released Hy4-preview, a 770-billion-parameter mixture-of-experts model that activates about 49 billion parameters per token. The company positions it for coding, office, scientific-research and financial-analysis tasks; weights and model materials are available through Hugging Face.

Why it matters: The release broadens the choice of high-capacity open-weight systems and makes sparse-model deployment economics more relevant. Vendor benchmark claims still require independent reproduction.

Impact: The active-parameter count does not make deployment small: memory, expert routing, interconnect and serving software remain substantial. Test licence terms, multilingual quality, tool-use safety and total cost on representative workloads.