Tencent

Tencent, an internet company founded in 1998 and headquartered in Shenzhen, China,¹ develops large language models under the Hunyuan (混元) name; its newest models are credited to the “Tencent Hy Team”.²³ The first open-weight language model on its tencent account on Hugging Face was Hunyuan-Large: a mixture-of-experts (MoE) Transformer, in which each token passes through only part of the network, with 389 billion total and 52 billion activated parameters and a context of up to 256K tokens Vendor-stated, described in a paper first posted on Nov 4, 2024.⁴⁵ Hugging Face’s spring 2026 report says Tencent increased its releases on the Hub eight- to ninefold in 2025.⁶

The Hy generation followed in 2026. Hy3 is a 295-billion-parameter MoE model with 21 billion active parameters and a 256K-token context, released under Apache 2.0 after what its card calls a Hy3 Preview launch “in late April” Vendor-stated.³ Hy4 preview, posted on Hugging Face on Aug 27, 2026, has 770 billion total parameters, 49 billion of them activated per token, with 256 routed experts and one shared expert in each MoE layer and a 1M-token context; it is also licensed under Apache 2.0 despite the preview label Vendor-stated.⁷⁵

On OpenRouter, an API service that routes requests to hundreds of models, Hy4 preview processed 8.31 trillion tokens in the seven days to Oct 12, 2026, fifth among all models and up 31%, and Hy3 was 13th with 2.78 trillion Measured; both rows are in Most-Used AI Models on OpenRouter by Weekly Tokens.⁸ On Hugging Face, Hy3 was Tencent’s most-downloaded general language model on Oct 13, 2026, with 578,301 downloads in the previous 30 days Measured, as Chinese Open-Weight Models: Downloads, Derivatives and Token Share records alongside eight other Chinese labs.⁵