Tencent's Hy4 adds vision and computer use to flagship model ahead of open-weight release

What Tencent shipped

Tencent released Hy4, the latest version of its flagship language model, on Nov. 4, 2026. According to The Daily Flare, the two headline additions are visual understanding and computer use: the model can interpret images, parse complicated charts and carry out mathematical reasoning on visual input, and it can operate command-line tools and graphical interfaces, extending it beyond purely text-based work.

TechNode described Hy4 as a 770-billion-parameter model. Tencent said the context window is one million tokens, matching Hy4 Preview, which the company introduced on Aug. 28. SiliconANGLE previously reported that the preview had 770 billion parameters, 49 billion of them active.

Tencent said it expanded both pre-training and post-training for this release, with reported gains in instruction following, long-context comprehension, extended agent tasks and deep research. It also said it corrected two flaws in the preview's thinking process: a habit of reasoning too much and of re-checking its work unnecessarily.

An agent for professional software

The company trained Hy4 to work with professional applications, naming Blender and computer-aided design tools, with the aim of planning and completing multi-step tasks. Game development is a particular focus. Hy4 is integrated with Unreal Engine, Unity and Godot to support complex game logic and multiplayer development, and Tencent reported a 19.1 percent improvement over the preview on its GameDev benchmark plus a 9.8 percent gain on GameLogic — though SiliconANGLE noted that the underlying baseline scores were not published.

Tencent also cited progress in front-end work, including business logic, state management and coordination among components, and said the model's long-context abilities are tuned for tasks such as web development, document summarization, visual design and building presentations.

Workplace push through WorkBuddy

Hy4 is built into WorkBuddy, Tencent's workplace AI tool, where the company reported an average score of 83.18 and a 96.15 percent pass rate on its evaluations, along with a first-place ranking in nine of 12 scenarios in the mainland Chinese market. Connected to WeCom, the model can draw on chats, email, documents, calendars and meeting records to track project history, surface unfinished tasks, flag meetings that need preparation and monitor deadlines. Collaborative features built on Hy4 are planned for WorkBuddy's forthcoming Projects feature.

Tencent also published a financial-analysis exercise run by its Hy team with Tencent Financial Technology: a blind evaluation covering 579 tasks based on real financial work, judged by 213 professionals, 153 of them from outside Tencent and about 70 percent with more than five years in the industry. The company said Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97, Kimi K3.1 Pro at 2.68 and GLM-5.3 at 2.68.

Availability and open weights

Hy4 is available now through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's AI Assistant and Ima, as well as third-party platforms including OpenRouter, Hermes and OpenCode. Tencent said model weights and system cards will follow in the coming days as part of its open-source development effort, naming cybersecurity and biological safety as areas of focus. A lower-cost Hy4 Flash version is also scheduled to follow.

Why it matters

The launch shows how quickly agentic capabilities — reading screens, operating software, sustaining multi-step work — are becoming standard features of flagship models rather than niche add-ons. Because Tencent intends to publish the weights, Hy4 offers developers and researchers an open-weight option with these abilities, plus game-engine integrations that most Western model makers have not prioritized.

The limits are spelled out largely by Tencent itself. The South China Morning Post reported that the company acknowledged Hy4 still trails the leading closed-source models on research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering. And the strongest numbers in the announcement — the benchmark deltas and the financial-analysis result — come from Tencent's own evaluations, so independent testing once the weights land will be the real test.