AI Momentum 107.6 +0.8%GitHub Pulse 5216 -10.2%Capital Flow $224.8B +3206.4%Hiring Heatmap 0.0 new
Advertisement
Open Source

Tencent's new AI model jumped into a major marketplace's top 10 at half price. The free version it promised hasn't shipped

A day after launch, Tencent's newest AI model, Hy4, was one of the most-used models on one of the largest AI marketplaces. Two things got it there. One was a half-price promotion. The other was a promise that Tencent would soon give the model away. The giveaway has not happened yet.

OpenRouter lets developers reach hundreds of AI models through a single account. Its public usage rankings for November 5 show Hy4 processing about 4 trillion tokens, the small chunks of text that AI usage is billed in. That was roughly triple the 1.3 trillion recorded the day before. It ranked 10th of 646 models, with about 2% of all traffic.

The gap between the free promise and the paid product matters because of how Chinese labs have won business. They publish "open-weight" models: the trained model files, which anyone can download, run on their own servers and modify. Buyers like that because it reduces dependence on one vendor and lets them keep sensitive data in-house. Tencent's August preview of Hy4 came out that way, under Apache 2.0. That standard license allows commercial use with no revenue limits. The finished Hy4 launched on November 4, and so far it is sold only as a service running on Tencent's own cloud.

What customers can actually get today

Tencent says it will publish Hy4's model files "in the coming days." It also promises a system card, a safety and technical report. It has given no date for either and has not said what license the files will carry. On November 5, the research firm Digital Applied found that Tencent's download page for the final model returned an access error and its code repository returned a not-found error. As of November 6, National Technology reported the files were still unreleased. Digital Applied also said Tencent had not confirmed the final model's size. The preview had 770 billion parameters, the internal settings that measure a model's scale.

Tencent's marketing is already using the word open. Its official social account calls Hy4 "firmly among the leading open models," according to Gigazine. Several outlets ran headlines calling it open source. Startup Fortune called the service-only launch a reversal of Tencent's usual approach. It warned that the move risks alienating the developers who made the previous model, Hy3, popular.

This also affects where customer data goes. OpenRouter lists exactly one company running Hy4: Tencent Cloud. So any business that sends work to Hy4 through OpenRouter is sending its data to Tencent's servers. Once the files are released, other hosting companies could run the model, and so could customers in their own data centers. Until then, no one can undercut Tencent on price or offer to keep data inside a particular country.

Price is doing much of the work

Tencent's list price is $0.834 per million tokens of input and $2.501 per million tokens of output. Tencent says that matches the preview's price. OpenRouter currently charges half that, with no stated end date. Tencent's own cloud store, TokenHub, offers 50% off through December. Its WorkBuddy and CodeBuddy work apps include free usage through November 15.

Early outside readings on value look favorable. Figures from the testing site Arena.ai, cited by Gigazine, put Hy4 at $0.55 per task. That is about 62% cheaper than Moonshot's Kimi K3.1 Pro, with scores within one percentage point. OpenRouter's listing also shows where demand comes from. The biggest single app sending Hy4 work is Claude Code, Anthropic's coding assistant. Developers can set it to use cheaper rival models in place of Anthropic's own.

Most performance claims still come from Tencent. On a test of operating a computer on its own, Tencent reports Hy4 scoring 78.9, close to Anthropic's Claude Opus 5.5 at 81.8. On a command-line coding test, Tencent puts Hy4 at 44.4, against 66.4 for Opus 5.5. Tencent concedes a "significant gap" to the best closed models. Trending Topics reported that Artificial Analysis, a widely followed independent tester, had not yet published results for Hy4.

Hy4's surge also leaves the market leader untouched. DeepSeek's V4.1 Flash, an open-weight model, held about 28% of OpenRouter traffic on the same day. That is roughly 14 times Hy4's share.

What buyers and investors should watch

First, the license. Several recent open releases have carried conditions such as revenue caps or limits on how the model can be used. If the final Hy4 keeps the preview's plain Apache 2.0 terms, any company can build on it. Any added condition would show up in the license text.

Second, whether usage holds when the discounts end. A 50% cut is a costly way to buy market share. Tencent also says a cheaper, faster version called Hy4 Flash is coming, again with no date.

Third, the safety report. Tencent says it removed material on attacking computer systems from Hy4's training data. National Technology noted that Moonshot took a similar step with Kimi K3.1 Pro. Enterprise risk teams will want the promised system card before they approve running the model in-house.

Tencent has shown that a strong model at half price can win developers within days. When the files arrive, and on what terms, will show whether "open" describes the product or just the promotion.