Simon Willison runs Tencent's Hy4: 770B total, 49B active, 1M context, 1.56TB of open weights
simonwillison.net·medium signal
Willison's August 29 post covers Tencent's Hy4 preview, a jump from Hy3's 295B total / 21B active / 256K context to 770B total, 49B active and a 1M token window, shipped on Hugging Face as a 1.56TB download. It exposes two reasoning effort levels, 'high' as default and 'no_think'. He noticed the reasoning trace is written in slightly truncated English and reads as the model deliberating, considering and then rejecting details like helmets and sunglasses in his SVG test.