Hunyuan 3.0

Tencent put the weights for Hunyuan 3.0 on Hugging Face under an Apache 2.0 licence, with a free API on OpenRouter through July 21. The model holds 295 billion parameters but lights up only about 21 billion of them for any given word — a sparse mixture-of-experts design where 8 of 192 sub-networks fire per token. That's the trick behind cheap Chinese open models like DeepSeek: it costs like a small model to run while carrying the knowledge of a large one.

It isn't beating the bigger models. It's losing by less than its size would predict — and doing it on the compute of a model a tenth as large.

Read the fine print in Tencent's own benchmark table. On the coding tests it actually shows head-to-head, Hunyuan 3.0 loses to the larger GLM-5.2 across the board. So the headline pitch — that it rivals models several times its size — isn't a claim to beat them. It's a claim to lose by less than its footprint would predict, for a fraction of the compute. That's a real argument, but it's about cost per token, not capability, and the two get quietly swapped in the marketing.

What makes this land beyond one more model drop is who shipped it. Tencent is the third serious Chinese lab — after DeepSeek and Alibaba's Qwen — to hand the West a frontier-adjacent model it can download and self-host, and it did so the same day the model went live inside WeChat and QQ Browser for a billion-plus users. Simon Willison ran his standing sanity check, an SVG of a pelican on a bicycle, and got a recognisable result — a clear step up from Tencent's spring preview. The weights are yours to inspect; treat the benchmark table as the sales deck it is.