On November 20, the weight files for Tencent's Hunyuan Hy4 appeared on the download pages of several major code and model platforms. No application form. No price.
The model card reads: 770 billion total parameters, a mixture-of-experts architecture, about 49 billion active, context up to one million tokens, and computer use opened for the first time.
One in fifteen
770 divided by 49 is 15.7.
For every 15.7 parameters in the repository, one takes part in a given forward pass.
That is what the architecture is for. Total parameters describe the size of the warehouse; active parameters describe how much is carried out per trip. A larger warehouse holds more without raising the cost of each trip. It also has to be rented whole. The goods you never move still occupy the shelf.
Both halves matter when 770B and 49B sit on the same card.
Big parameters do not make a call expensive. They make the memory bill large.
What the 49B pays for
The 49 billion figure sets the compute per token. It is 15 times smaller than 770 billion, and about the same as a dense 49B model.
Whether a call is cheap depends on 49B. Whether it fits on your machine depends on 770B.
Those two questions get collapsed into one complaint, that the model is too big, and they belong to different budgets: one to the electricity bill, one to the server room.
Had Hy4 been dense, every forward pass would touch all 770 billion parameters. At 49 billion the arithmetic drops to roughly one-fifteenth. That is an order of magnitude, not a rounding. Two models both labelled 770B can sit in different cost classes depending on sparsity, and the letter B on a model card says nothing about which class you are in.
The parameter count is written for the people who download. The active count is written for the people who call.
What is being given away
Open source here means weights. A downloader gets a set of files that can run locally, be modified, fine-tuned or redistributed.
A service is a different object.
The Cyberspace Administration of China counts something else. In September and October 2026, 133 generative AI services completed filing, 3 of them on-device for phones, 1 an anthropomorphic interaction product, and 1 an update to an existing filing.
That is a list of services offered to the public. A weight download is not on it.
Ten thousand downloads can produce zero launched services.
Two services a day
133 services across the 61 days of September and October is 2.2 a day.
That is the rate of filed services, not of models. One model can carry several services. One service can swap models.
The numbers do not line up because the units do not. 133 belongs to products that face users. 770B belongs to files that run on hardware that someone already owns.
One path starts with a filing, then users, then usage. The other starts with a machine, and a machine does not need a user to be switched on.
The part that can click
Of the capabilities listed for Hy4, computer use is the one worth reading twice. The model can read an interface, press buttons, fill fields, scroll a page.
Million-token context is the other.
Together they point at one kind of work: hand over a whole codebase or an admin console and let the model find its own way around. Long inputs, small actions, both constraints tight at once. A chat model needs neither. A model meant to do the work needs both.
One reads long. One can reach the keyboard.
Computer use carries a second implication. If it can press a button, it can change something. What it changes, which file or which account, depends on the permissions granted by whoever installed it. Open weights ship capability, not authority. Authority stays with the deployment, and behind every button the model is allowed to touch.
What open weights buy
Tencent is not charging for the download. It is collecting something else.
The default.
Once developers pull the weights, inference frameworks, quantisation tools, fine-tuning scripts and deployment docs grow around them. That growth is sticky. When a company eventually buys, the first row of the technical evaluation is often the model its team already runs.
That is habit, not contract. Habit is harder to renegotiate.
A model installed on someone's own hardware gets a next call. A model waiting behind an API has to clear a procurement process first.
Private deployment behind a corporate firewall, continued training on industry data, a model baked into a hardware box: all three begin with a file, not an endpoint. Weights reach those places. An API does not.
The bill nobody can calculate
Tencent has not disclosed what training cost.
So the return on this release cannot be computed. What can be computed is the threshold: the download is free, and deployment costs one card that can hold 770 billion parameters, or a set of them.
The threshold decides who can use it.
A 770B checkpoint handed to a team of twelve will probably not run. Handed to a company with a cluster, it serves traffic the same day. Open weights flatten the price. They do not flatten compute.
The marginal cost sits in bandwidth and storage, not in training, which is paid once and sunk. Distribution is paid per download, and Tencent pays it.
After November 20, the files are simply there. 770B, 49B, a million tokens of context, computer use. The model card lists them. The download button sits underneath.
No contract. No price list.
The only gate left is memory: all of 770B has to fit, and only 49B is computed at a time. The download page hands over the file. Everything after that is yours to solve.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
