BLOG / November 6, 2026
November 4: Tencent's Hunyuan... Hy4 It launched products such as WorkBuddy; on November 5, the company announced that it had experienced a surge in requests and had urgently scaled up its inference cluster by about tenfold; on November 6, China News Network again highlighted Agent Arena’s “No. 1 in open source” status and Artificial Analysis’s “No. 2 in China” ranking in its headlines.
It’s certainly buzzing. But there’s another, more jarring fact from the same week: the official copyright notice still reads “to be open-sourced soon,” and what’s available for download on Hugging Face is still Hy4-preview. If you interpret “open-source models first” as “you can take the weights home today,” you’ll run into trouble.
The demand for a tenfold scaling solution has come flooding in; the Weighted Schedule Proof is still just a roadmap—not a download button.
Hit the daily limit on the first day; capacity increased tenfold
According to an announcement by Tencent Hunyuan and a report by Yicai: After Hy4 was first launched on WorkBuddy on November 4, a large number of users reached their usage limits on the very first day. The team urgently scaled up the Hy4 inference cluster by approximately tenfold and increased the free trial quota, while also stating that it would continue to dynamically allocate resources.
The guidelines are also quite clear: there may still be limits on total high-end computing power and peak concurrent usage. Transition measures include switching to Hy3 (the free trial for WorkBuddy has been extended until 11:59 PM on November 15, 2026) and, for non-urgent needs, using the service during off-peak evening hours. This is not an “unlimited free-for-all,” but rather a policy of “prioritizing user experience first, then managing wait times.”
Official pricing guidelines remain unchanged: 6 yuan per million tokens for input, 18 yuan per million tokens for output, and 0.3 yuan for a cache hit; the 50% discount on TokenHub’s published rates has been extended through December 31, 2026. Long-context and multi-step tool calls by Agents can quickly burn through tokens—once the discount window closes, the billing narrative will change.
Agent Rankings and the Term “Open Source”
A November 6 report by China News Network stated: In Artificial Analysis’s latest AI Index, Hy4 ranked second among Chinese models, trailing only Kimi K3.1 Pro and outperforming Qwen3.8 Max, DeepSeek V4.1 Flash, and others; The same article cited a cost efficiency figure of approximately $0.78 per task. On Agent Arena, Hy4 ranks around 13th overall, is the top open-source model, and lies on the Pareto Frontier.
These are excerpts from review platforms and media reports; they are not original content from this site. When reviewing the rankings, keep the following distinctions in mind: the overall intelligence index ≠ the success rate of long-running agents in your business; ranking first on an open-source list ≠ publicly disclosed weightings.

At the same time, Tencent’s official website emphasized that, compared to the Hy4 preview, the official version expands the scale of pre-training and post-training, introduces visual understanding for the first time, and enhances instruction following, long-text processing, and long-range agent capabilities; it also supports “Computer Use” for the first time and provides specialized training for professional software such as Blender and CAD. An internal evaluation by financial experts (213 participants, 579 tasks) ranked the Hy4 open-source model first overall—though this is also labeled as an official/collaborative evaluation, not a standardized external competition.
The weighting is still "coming soon"
The original text on Tencent’s official website states clearly: “Hy4 will open-source the model weights and release model system cards in the near future.” As of the time of this article’s verification, the available downloads listed on the Hugging Face organization page are still tencent/Hy4-preview(Approx. 780B parameters, Apache-2.0), but no complete weight package has been found that corresponds one-to-one with the capabilities claimed for the official release.


English-language tech reviews have also pointed out the same inconsistency: the word “open” in the title and “schedule” in the body are written separately, which could easily lead readers to mistake the preview repository for the official release. On the product side, it can already be called from entry points such as WorkBuddy, TokenHub, and OpenRouter; on the repository side, however, “Agent open-source first” has not yet been translated into “the same set of weights that can be reproduced locally.”
Back to the product roadmap: Hy4 is already available through entry points such as Tencent Cloud TokenHub, WorkBuddy/CodeBuddy, Yuanbao, the WeCom smart assistant “Dayuan,” and ima, and its API has been integrated with platforms like OpenRouter; integration with QQ, gaming, and content-related products is still ongoing. In other words, Tencent is first rolling out “functional agent entry points” and then using scalability and free trials to offset the traffic surge during the first week.
Security narratives were also included in the press release: on the training side, the emphasis was on incorporating defensive cybersecurity data and excluding offensive exploitation content; on the biometrics side, it was noted that while participants had a strong foundation of basic knowledge, their ability to execute experimental protocols was limited. These are the official self-assessment findings; external independent red team reports remain to be seen.
The Hy4 Flash is billed as a successor model that builds on its predecessor’s capabilities while reducing task costs. If you’re more concerned with your bill than the charts, the Flash and the expiration dates of current discounts (TokenHub at 50% off through December 31, WorkBuddy free through November 15) will influence your actions next week more than the phrase “open source first.”
Looking at the three timelines side by side makes things clearer: the release and product integration on the 4th, the scaling up and quota announcement on the 5th, and the growing buzz around the rankings on the 6th. Media headlines love to weld “tenfold” and “number one” together, but readers need to break them down and evaluate them separately—capacity expansion is an operational fact, the rankings are a snapshot of performance evaluations, and weighting is a data-repository fact. All three can be true at the same time, yet headlines can twist them into a single false causal statement.
For those building Agent applications, short-term actions are more like “testing the entry point, monitoring traffic throttling, and noting discount expiration dates” rather than “cloning the weighting model today to build a private version.” Percentage improvements in Computer Use, financial beta testing, and GameDev should all be labeled according to official or partner evaluation metrics; until independent replication is available, they are better suited as topics to watch rather than as acceptance criteria in procurement contracts.
Don’t overlook the impact of the preview version either: English-language comments point out that Hy4-preview is still contributing a significant amount of tokens on platforms like OpenRouter. While the narrative around the official version’s capabilities is more impressive, the version available for download is still the old checkpoint—which means that community benchmarks, quantification, and secondary fine-tuning will, for the time being, have to revolve around the preview version, which may not be isomorphic to the “Hy4” in WorkBuddy.
A quick takeaway on choosing the right approach: If you want to experience the Long-Range Agent, use the official product and API; if you want auditable, open-source reproducibility, wait for the system cards and the official weight repository—don’t simply translate Agent Arena rankings directly into Git LFS files.
Our assessment:Hy4: The real news this week isn’t yet another headline about “yet another flagship open-source project,” but rather the fact that “product-side calls being overwhelmed” and “weights still not implemented” appeared side by side. It’s fine to track the rankings, but for the download button, you’ll have to wait for the system card and the official repository; the tenfold expansion indicates that the Agent entry point has won the attention battle for now, but it doesn’t mean the window for local deployment has opened yet.
Reference Links
- Tencent Official Website: Hun Yuan Releases Hy4
- Yicai: Hy4 Calls Surge; WorkBuddy Scales Up Tenfold
- China News Network: Tenfold Scaling and Agent Arena’s First Open-Source Release
- Hugging Face: tencent/Hy4-preview
- GitHub: Tencent-Hunyuan/Hy4-preview


Comment List (0):
Load More Comments Loading...