On August 28, Tencent released Hy4 preview — a 770-billion-parameter mixture-of-experts model that activates only 49 billion parameters per request — and open-sourced it. The pricing: $0.834 per million input tokens, $2.501 per million output. That’s roughly 72% cheaper than Kimi K3, the open-weight incumbent from Moonshot AI, which lists at $2.55 input and $12.75 output on OpenRouter. For two weeks, Hy4 is free on Tencent’s WorkBuddy and CodeBuddy platforms.

The reaction from the open-source faithful was predictable: another win for the commons, another Chinese lab proving that open models can match closed ones, another nail in the coffin of the “AI is too dangerous to open-source” argument. Hacker News lit up. The benchmarks got screenshotted.

All of that misses what’s actually happening.

Tencent Is Not a Lab. It’s a Conglomerate.

Tencent doesn’t need Hy4 to make money. Tencent runs WeChat, the most successful super-app on earth. It runs one of the world’s largest gaming businesses. It runs Tencent Cloud, which competes with Alibaba Cloud and Huawei Cloud for enterprise customers across Asia. It runs a fintech arm that processes payments for over a billion people.

When a company like that open-sources a frontier-tier model, the question isn’t “why are they being generous?” The question is “what are they selling?”

The answer is on the same page as the release announcement: Hy4 is available through Tencent Cloud TokenHub. It’s free on WorkBuddy and CodeBuddy for two weeks. Free access to the previous model, Hy3, has been extended to September 30. This is a promotional calendar, not a manifesto.

The model is the toaster. The bank account is the cloud subscription.

The Marginal Cost of Frontier AI Is Collapsing

Here’s the number that should make everyone in the AI industry pause: a 770-billion-parameter model — a model that, two years ago, would have been the most advanced system on the planet — is now being given away as a two-week free trial to drive signups for a productivity app.

Not a startup burning venture capital to buy market share. A profitable conglomerate using a frontier model as a customer acquisition cost.

One developer I spoke to on a Discord server for AI infrastructure engineers put it bluntly: “Two years ago I was rationing my GPT-4 API calls. Now Tencent is handing out a 770B model like it’s a branded USB stick at a trade show.”

The implication is uncomfortable for a lot of people. If the marginal cost of producing and serving a frontier-tier model has fallen far enough that a conglomerate can use it as a promotional item, then the “AI arms race” framing — the idea that whoever has the best model wins — is already obsolete. What matters now is distribution, integration, and switching costs. Tencent has all three.

The Open-Source Ideologues Need to Update Their Priors

There’s a certain type of person who treats every open-source model release as a moral victory. The narrative goes: open models democratize AI, closed labs are rent-seekers, and the future belongs to the commons.

Tencent’s Hy4 release should complicate that story. This isn’t a collective of researchers sharing weights out of principle. It’s a publicly traded company with a market cap in the hundreds of billions, using open-source as a distribution strategy. The weights are free because the lock-in is downstream — in the cloud credits, the enterprise seats, the WeChat integrations, the API calls that eventually migrate from OpenRouter to Tencent Cloud TokenHub.

That’s not a criticism. It’s a description. And it’s a more honest description than either the open-source romantics or the national-security hawks are offering.

The hawks, for their part, will see a Chinese conglomerate undercutting Western labs on price and reach for the usual language about industrial policy and strategic competition. They’re not wrong that Tencent benefits from a domestic ecosystem that Western companies can’t replicate. But they’re missing the more mundane truth: Tencent is doing what every conglomerate does when a new technology gets cheap enough. It’s using the technology as a loss leader to sell something else.

The open-source AI movement didn’t win a moral victory on August 28. It got a pricing signal. Frontier AI is now cheap enough to be a promotional item. That’s a bigger story than any benchmark table.

Sources