Tencent's 770B Hy4 weights make model spend a memory-capacity question

Sparsity is the first number I would put on the whiteboard: 49 billion active out of 770 billion total ^[1] is 6.4% of the weights doing work on any given token ^[11], against 7.1% for July's 295B/21B Hy3 ^[2][12], so the model got 2.6 times heavier to store ^[10] while the per-token compute path widened only 2.3 times ^[8].

The file size is where the buy-side arithmetic actually lives. Divide 1.56TB by 770 billion parameters and you get roughly 2.03 gigabytes per billion, which is two bytes a parameter, and Hy3's 598GB over 295 billion is the same 2.03 ^[13], so nobody has quantized anything on your behalf and the memory problem is yours to solve. At 80GB per accelerator, 1.56TB is twenty devices before the first token comes out ^[14]. That is the honest entry price of the build column.

Free weights still cap what a vendor can charge for text inference, though, because the buyer's alternative was never only self-hosting. Simon Willison ran his pelican-on-a-bicycle prompt through OpenRouter at the default high reasoning setting ^[4], which means somebody else had already bought the memory and was reselling access to it, and a reseller of downloadable weights competes on margin rather than on capability. I do not have those per-token prices in front of me. It is the first number I would go and get.

Tencent has put 1.56TB on Hugging Face ^[1], and that choice is its own statement: it cannot now sell access to this generation's text capability, so it is buying distribution and default-choice status with revenue it will not collect a token at a time.

The reasoning dial is the finding here, because there isn't one. The chat template accepts only "high" and "no_think" ^[3], so a team that wants a third of the thinking budget has no setting to ask for it, and hidden reasoning tokens are precisely the line item that breaks a monthly estimate. Willison noticed the trace writing in clipped, half-grammatical English, which he reads as perfect grammar not being worth the tokens ^[6]; that is a vendor compressing its own cost, and it implies the thinking runs long enough to be worth compressing.

This is probably wrong, but the direction I would bet on is that the marginal price of frontier-adjacent text tokens converges on hosting cost plus a thin margin. What would prove it wrong: hosted Hy4 prices arriving close to closed-model text pricing, which would say accelerator memory rather than model access is the genuinely scarce good.