AI Momentum 110.7 +2.9%GitHub Pulse 5178 -6.1%Capital Flow $204.6B +12563.4%Hiring Heatmap 0.0 -100.0%
Advertisement
Open Source

Tencent gives away its top AI model, but big companies reselling it now need Tencent's permission

Tencent on November 20 released the weights of Hy4, its flagship AI model. Weights are the downloadable files that let anyone run a model on their own computers. The release came 16 days after Tencent promised it "in the coming days."

Two days earlier, Tencent did the same for Hy4 flash, a cheaper sibling. Flash is now the most-used model by daily volume on OpenRouter, a popular marketplace that routes developers' requests to hundreds of AI models.

The bigger change for buyers is in the license. Hy4's August preview used Apache 2.0, a standard open-source license with essentially no commercial limits. Both new releases instead use a custom "Tencent Hy License." It is based on Apache 2.0, but adds a clause aimed at companies that sell access to the model as an online service.

What the new license changes

Under the clause, a company offering the model to outside customers as a hosted service needs a separate written license from Tencent in two cases. One is if its worldwide revenue topped $50 million in any 12-month period. The other is if the product has more than 100 million monthly users.

The rule also covers modified copies. That includes versions retrained for a specific task, compressed versions and models merged with others.

Purely internal use is exempt, as long as neither the model nor its output reaches outsiders. So are products that use the model for a narrow feature rather than reselling general access. Individual developers and smaller firms can still build on it and charge for what they make.

That leaves the condition falling mostly on businesses that would compete with Tencent Cloud for paying customers: other cloud providers and AI hosting firms. The license sets no price for the separate deal. It lists only a partner email address. The text contains no country restrictions.

A popular model with one seller

The new terms arrive just as Tencent's cheap model is drawing heavy traffic. OpenRouter's own rankings, with data through November 19, show Hy4 flash handling about 22.6 trillion tokens in a day. Tokens are the word fragments AI use is billed in. That was more than twice DeepSeek's most-used model, in second place.

A separate tracker, Tokenmaxxing, put Hy4 flash at 23.8 trillion tokens on November 18, also first. TechNode reported it ranked first on OpenRouter's daily chart on November 17, the day after Tencent launched it as a paid online service.

Every one of those requests runs through Tencent. OpenRouter lists a single provider for Hy4 flash: Tencent Cloud. By Neurohive's count, DeepSeek's competing V4.1 Flash is served by more than 30 providers.

Free weights normally invite rival hosts like those, and the competition pushes prices down. The new license makes the largest of them negotiate with Tencent first.

Price is the main draw. Tencent lists Hy4 flash at about $0.07 per million input tokens and $0.28 per million output tokens. OpenRouter shows a further 50% discount on top. Tencent says a typical task costs one-twentieth of what it costs on the full Hy4.

Tencent also claims Hy4 flash processed more than 50 trillion tokens on its first day. No one outside the company has checked that figure.

What independent tests show

Cheap does not mean top quality. Neurohive and Hermes both reported that Hy4 flash scored about 51% on Vals AI's independent index. That put it near the bottom of roughly 40 ranked models.

On Neurohive's figures, that is about 8 points behind Anthropic's Claude Haiku 5.5. Tencent's own blind test with 85 users had scored the two as a near tie. Neurohive also reported OpenRouter speeds of about 70 tokens a second, roughly half of what Tencent claims.

The full Hy4 has no independent scores yet; its benchmark results are Tencent's own. It is also expensive to run: Tencent's setup guide spreads it across eight AI chips working together.

For buyers, the result is mixed. Companies that run Hy4 inside their own walls gain a capable free option with few conditions. Cloud providers and AI resellers now face a negotiation Tencent did not require in August.

Downloads are still modest. The Flash page on Hugging Face, the main site for sharing AI models, showed about 2,000 downloads in the past month. The August preview showed about 11,000. The thing to watch is whether outside hosts sign deals with Tencent, or stay away.

Written by AI from the public sources linked below.