Tencent's Hy4 Officially Launches: Vision and Computer Use Added, Offensive Cybersecurity Content Removed From Training
Tencent released Hy4 on November 4, 2026: a 770-billion-parameter mixture-of-experts model with vision understanding and computer use for the first time, trained with offensive cybersecurity content removed and only defensive content retained. The weights and system card are still to be released. API pricing matches the preview version, with Tencent Cloud's TokenHub at half price through December 31.
- 01THE ANNOUNCEMENT — Reuters reported on November 4, 2026 that Tencent released Hy4, removing exploit development, reverse engineering and penetration testing from training and keeping only defensive cybersecurity content.
- 02THE CONFIRMATION — TechNode's independent same-day report confirmed the launch: a 770-billion-parameter mixture-of-experts model with vision understanding and computer use for the first time, live in Tencent products and on OpenRouter.
- 03THE WATCHPOINT — The model weights and system card Tencent releases later, plus Hy4 Flash's specifications and pricing, will determine whether outsiders can independently verify the safety removals and real capability.
WHAT HAPPENEDWhat Happened
Reuters reported on November 4, 2026 that Tencent released Hy4 on Wednesday, and that the company said the model was trained with only defensive cybersecurity content included, to reduce the risk of the model abetting cyberattacks. In its statement, Tencent said it recognizes the potential security risks posed by advanced cyber and biological capabilities and is committed to training, open-sourcing and releasing models in a more responsible manner.
The key qualification in the same Reuters article: Tencent filtered exploit development, reverse engineering and penetration testing out of the training data to limit Hy4's offensive capability. Reuters also noted that Moonshot AI applied similar limits on cyber-offense capability to its latest model last month, and that Chinese vendors, like US companies, are guarding against loss-of-control risks.
TechNode's independent report of November 4, 2026 (09:15 UTC): Hy4 launched on November 4 with new vision understanding and computer-use capability — it can interpret charts and documents, reason across multiple images, and operate command-line tools and graphical interfaces, with targeted training for software such as Blender and CAD. The 770-billion-parameter model is live in Tencent's own products and on third-party platforms such as OpenRouter. Tencent plans to release the weights and system card later, and a low-cost Hy4 Flash is also planned.
Deployment details from the same TechNode article: Hy4 is integrated into WorkBuddy and CodeBuddy, with a limited free quota from November 4 to November 15. Standard API pricing matches the Hy4 preview — $0.834 per million input tokens and $2.501 per million output tokens — with Tencent Cloud's TokenHub at half price through December 31. Development focused on office tasks, financial analysis and game development, with product teams taking part in training and real-workflow evaluation.
WHY IT MATTERSDeleting offensive data is the new normal for open-source releases
Hy4's release method deserves more attention than its parameter count. Within a month, Moonshot first applied cyber-offense limits, and Tencent has now followed with the same approach: before open-sourcing, remove exploit development, reverse engineering and penetration testing from the training data, keeping only defensive content such as vulnerability repair and proof-of-concept generation.
Once this becomes the default among leading vendors, every open-source release gains a verifiable new threshold: not whether the model can perform offensive tasks, but what the vendor proactively removed and kept before release, and whether the removals are documented in the system card. Hy4's system card has not been published, so that verification is still missing its final piece.
WHY IT MATTERSVision plus computer use fills Hunyuan's gap
Hy4 is the first model in the Hunyuan line to combine vision understanding with computer use: reading images and tables, reasoning across images, then directly operating command lines and graphical interfaces to run multi-application tasks. Office work, financial analysis and game development are listed as development priorities, and product teams took part directly in training and evaluation — a positioning aimed not purely at benchmark rankings but at real workflows.
But the state of play on release day must be stated plainly: what launched is a service, not downloadable weights. Third-party platforms and Tencent's own products can call Hy4, while the weights and system card remain to be released later. Until the weights are genuinely public, outsiders cannot independently reproduce the model or verify that Tencent's safety removals work as described.
SIGNAL JUDGMENT — ANALYSISThe signal is the release method, not the parameter count
Tokenta Signal's judgment: the signal in Hy4 is not its 770 billion parameters, but Tencent choosing the same path as Moonshot AI — safety removals before open-sourcing. Parameters and capabilities going live are only the starting point. Once the weights and system card are published, whether outsiders can verify the effect of the removals is the measure of whether this release route holds up.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
Reuters' November 4, 2026 report was opened and verified in full. Core facts (Hy4 released Wednesday, only defensive cybersecurity content retained, exploit development, reverse engineering and penetration testing filtered out, Moonshot AI's similar restriction last month) agree with TechNode's independent same-day report (770 billion parameters, vision and computer use, product launches and pricing basis). Both were opened.
PRIMARY SOURCESPrimary Sources
- Reuters report
https://www.reuters.com/legal/litigation/tencent-rolls-out-ai-model-trained-with-limited-cybersecurity-data-amid-hacking-2026-11-04/ - TechNode independent report
https://technode.com/2026/11/04/tencent-launches-hy4-with-vision-and-computer-use-capabilities/
RELATEDRelated Intelligence
- Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall Behind
- An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in Lean
- IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by Nighthawk
- Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 Outlook