Evening Brief

Tencent unveils Hy4 as Moonshot posts Kimi K3.1 Pro weights

Tencent says Hy4 ranks among top open models, with weights coming in days. Moonshot posted Kimi K3.1 Pro’s weights, with limits for big commercial users.

  1. Tencent — Tencent launches Hy4, joining the ranks of the world’s leading open-source AI models (Nov. 4, 2026)
  2. Hugging Face — moonshotai/Kimi-K3.1-Pro model card
  3. Hugging Face — Kimi K3.1 Pro License
A long, low white office building with vertical fins behind a row of trees and an empty road under a clear blue sky
Tencent’s Beijing office in Zhongguancun Software Park. Photo: N509FZ / Wikimedia Commons (CC BY-SA 4.0, cropped)

Tencent released Hy4, a large language model that can now read images and operate a computer, and says it ranks among the world’s leading open-source models. But the weights, the files that let anyone run the model, are not out yet. Tencent “plans to release model weights and system cards in the coming days,” its announcement says.

Moonshot AI, another Chinese lab, has already posted its weights. Kimi K3.1 Pro went up on Hugging Face Sunday night, the site’s records show, under a license with strings for big commercial users. It is a follow-up. When we reported the model’s launch on Oct. 23, Moonshot said open weights would follow by Nov. 4. They arrived ahead of that date.

Tencent’s Hy4

Tencent says Hy4 adds visual understanding and computer use for the first time. It can work in command-line and graphical interfaces, and was trained on professional software such as Blender and computer-aided design tools, the company says.

The scores are Tencent’s own. In a blind test of 579 financial tasks judged by 213 finance professionals, Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each, Tencent says. It says internal tests cut the hallucination rate to 7.2% from 12.95% in the Hy4 Preview.

On safety, Tencent says it left offensive exploit material out of training. Its own security labs found “no significant concerns” about offensive hacking ability in the areas they tested. Tencent says Hy4’s skill at open-ended experimental design is “more limited compared to frontier models.”

Hy4 runs now in Tencent apps and its TokenHub cloud service, with access through OpenRouter and other platforms. The price is about 90 cents per million input tokens and $2.70 per million output tokens, unchanged from the preview. TokenHub offers 50% off through Dec. 31.

Tencent’s post gives no parameter count, context length or license.

Moonshot’s Kimi K3.1 Pro

Kimi K3.1 Pro is a mixture-of-experts model, which routes each token to a small set of specialist subnetworks. It has 2.8 trillion parameters, with 104 billion active per token, and reads text and images, its model card says. It handles about 1 million tokens of context. Moonshot also offers it through its own application programming interface.

Its results are mixed by Moonshot’s own tests. It scores 65.7 on Terminal-Bench 4.0, a test of command-line tasks, against 66.4 for Anthropic’s Claude Opus 5.5. On Program Bench it scores 31.5 to Opus 5.5’s 65.0.

The license

Moonshot’s license lets anyone use, change and sell the model, with two conditions. Companies that sell model access as a service, with revenue above “20 million US dollars” over “any consecutive 12 months,” must sign “a separate agreement with Moonshot AI” before commercial use. Products with “more than 100 million monthly active users, or more than 20 million US dollars” in monthly revenue must “prominently” display “Kimi K3.1-Pro.” Internal use is exempt.

Our read: “Open” is doing a lot of work in both announcements. Tencent calls Hy4 a leading open-source model before anyone outside can download it, test it or read its license. That claim should wait for the files. Moonshot delivered weights before the date it promised, which deserves credit. But its license is not a plain open-source grant. A model-hosting business with more than $20 million in revenue over 12 months needs Moonshot’s sign-off, and a hit app must carry Moonshot’s name. Those terms are fair for a company protecting its work, and they leave most developers free. Developers should still read them before building a product on the model. Both companies lean on their own benchmarks, and Tencent’s headline scores come from its own products and partners. Chinese labs keep releasing strong models with weights. That is real pressure on paid rivals. Outside tests will show how strong.

Also today

OpenAI agrees to buy Pierre, maker of Code Storage, to bolster Codex. Conductor, an app for running coding agents, joins Meta. The New York Times and three papers sue Brave over AI scraping. How voters treated data center ballot measures. Salesforce agrees to buy Cresta for voice agents. CrowdStrike’s AI kill switch.

What’s next

Tencent says Hy4 weights and system cards come “in the coming days” and a cheaper Hy4 Flash “soon.” Its free WorkBuddy and CodeBuddy offer runs through Nov. 15.

Sources: Tencent, Hugging Face model card, Kimi K3.1 Pro License

Edition Nov. 4, 2026 (5pm) · Archive shelf