Tencent open sourced a 770B model and on its own blind test it loses about four times in ten

Tencent published Hy4 preview on 28 August under Apache 2.0, weights included, commercial use permitted. It is a mixture of experts, 770B total parameters, about 49B active per token, with context stated above 1M.

The number that stuck with me came from Tencent's own evaluation. 163 internal experts blind rated 203 engineering tasks. Against Kimi K3, its nearest rival, Hy4 won 51.2%, tied 7.9% and lost 40.9%. Four losses in ten, on a test the vendor ran itself, and Tencent's own word for the gap is slightly ahead. There is still no Artificial Analysis listing, and the Arena WebDev placement came out of AutoEval, where a reward model casts the votes rather than people, so I would not read either as independent confirmation.

Tencent Cloud lists Hy4 preview at $0.834 per million tokens in and $2.501 out. Tencent's own Hy3 endpoint on OpenRouter runs $0.132 in and $0.528 out at the standard rate, so the flagship is roughly six times the previous generation on input.

The part that reaches the filings is the bill behind it. Tencent reported the June quarter on 12 August: revenue RMB 204.8bn, up 11%, non-IFRS net profit RMB 68.4bn, up 9%, and capital expenditure of RMB 52.8bn against RMB 19.1bn a year earlier. Free cash flow was negative RMB 13.8bn, which the company puts at positive RMB 37.6bn excluding prepayments for compute. Giving the weights away sits inside that, and one quarter will not tell me whether it pulls paid volume forward or hands it over.

My exposure here is indexed rather than direct. Tencent was 8.43% of CNQQ on 31 August, against a rule that stops any single position at 10%, so about a point and a half separates the two. CQQQ is the nearer comparison on scope, since its index admits mainland listings at 25% of their full weight.