/
770 Billion Parameters 49 Billion Active 1 Million Context Native Visual Understanding Computer Use CLI GUI
Tencent launched Hy4 on November 4 2026 a 770 billion parameter mixture of experts model with 49 billion active parameters and 1 million token context. First Hy generation with native visual understanding and computer use for CLI and GUI. Hallucinations down to 5.45 percent. Weights coming soon.
Tencent launched Hy4 on November 4 2026. It is a 770 billion parameter mixture of experts model with 49 billion active parameters per token and a 1 million token context window. This is the first Hy generation to introduce native visual understanding and computer use capabilities for both command line and graphical interfaces. The model has been trained on professional software including Blender and CAD tools to handle long horizon agentic tasks.
Hy4 ranks first among open weight models on a finance evaluation scoring 3.04 across 579 financial tasks with a 12.03 percentage point gain over Hy4 Preview on iRaB. Hallucination rates fell from 15.76 percent to 5.45 percent and speculative errors dropped from 23.1 percent to 5.6 percent.
Weights and system cards are not yet available. Tencent says they will release them in the coming days. Until then Hy4 is accessible through Tencent Cloud TokenHub WorkBuddy CodeBuddy Yuanbao WeCom AI Assistant and ima. Third party API access is available through OpenRouter Hermes and OpenCode. Pricing is 0.90 dollars per million input tokens 2.70 dollars per million output tokens and 0.045 dollars per million cached input tokens. Hy4 Flash is scheduled for release soon with greater cost efficiency.
| Aspect | Hy4 | Hy4 Preview |
|---|---|---|
| Release date | November 4 2026 | August 28 2026 |
| Total parameters | 770 billion | not disclosed |
| Active parameters | 49 billion | not disclosed |
| Context window | 1 million tokens | 1 million tokens |
| Visual understanding | Yes first time | No |
| Computer use | Yes CLI and GUI | No |
| Hallucination rate | 5.45 percent | 15.76 percent |
| Speculative errors | 5.6 percent | 23.1 percent |
| Finance evaluation score | 3.04 across 579 tasks | lower |
| iRaB improvement | plus 12.03 percentage points | baseline |
| Weights available | Coming soon | Coming soon |
| Input price per million | 0.90 dollars | 0.90 dollars |
| Output price per million | 2.70 dollars | 2.70 dollars |
| Cached input price per million | 0.045 dollars | 0.045 dollars |
A frontier large language model announced November 4 2026 by Tencent Holdings Ltd. Runs on Tencent Cloud TokenHub and integrated into WorkBuddy CodeBuddy Yuanbao WeCom AI Assistant ima and more. Third party API via OpenRouter Hermes and OpenCode. MoE architecture with 770 billion total parameters and 49 billion active per token. 1 million token context window. Native visual understanding for fine grained image processing complex chart analysis and mathematical visual reasoning. Computer use capabilities for CLI and GUI including Blender CAD and game engines. Trained for long horizon agentic tasks.
Tencent Cloud account for TokenHub access. WorkBuddy or CodeBuddy subscription for free access November 4 to 15. Third party API key from OpenRouter Hermes or OpenCode for API access. Weights not yet released so self hosting is not possible today.
Not open weights yet. Not self hostable today. Not a standalone downloadable model. Not available on GitHub or Hugging Face as weights yet. Tencent explicitly states weights and system cards will arrive in the coming days.
Hy4 uses a mixture of experts architecture. 770 billion total parameters with 49 billion activated per token. This activates a subset of experts for each input which reduces compute cost while maintaining model capacity. Tencent has not published the exact number of experts or the routing mechanism details.
This is the first Hy generation with visual understanding. Capabilities include fine grained image processing complex chart analysis and mathematical visual reasoning. The model can read charts formulas and visual data natively without separate vision encoders or multimodal adapters bolted on after training.
Hy4 can interact with both command line interfaces and graphical user interfaces. Specialized training on professional software such as Blender and CAD enhances its ability to plan and execute long horizon sophisticated tasks. This puts Hy4 in the same category as models with computer use or agentic desktop automation capabilities.
1 million token context window matches the range of current open weights frontier models from European and Chinese developers including Mistral AI Beijing Moonshot Technology and Z.AI. Optimized for long context learning to understand complex information including web development document summarization visual creative tasks document presentation and more.
Hy4 ranked first among open weight models on a finance evaluation. Scored 3.04 across 579 financial tasks in blind evaluation. Showed plus 12.03 percentage points improvement on iRaB over Hy4 Preview. This is Tencent's claimed strongest external validation.
Hallucination rate fell from 15.76 percent in Hy4 Preview to 5.45 percent in Hy4. Speculative errors dropped from 23.1 percent to 5.6 percent. These are self reported figures from Tencent.
No third party benchmark results available yet. No public technical report. No system card yet. No weights to independently verify. All performance claims are Tencent self reported until weights release and independent evaluation.
Hy4 maintains the same pricing as Hy4 Preview despite significant performance improvements. Hy4 Flash is scheduled for release soon with greater cost efficiency for users who prioritize high value applications.
Tencent Cloud TokenHub WorkBuddy CodeBuddy Yuanbao WeCom AI Assistant ima. Third party platforms OpenRouter Hermes OpenCode. Integration efforts underway for QQ Honor of Kings Game of Peace WeRead and Tencent News.
From November 4 to 15 WorkBuddy and CodeBuddy will be available for free. Hy4 available on Tencent Cloud TokenHub at 50 percent off standard pricing through December 31 2026.
| Tier | Input per million | Output per million | Cached input per million |
|---|---|---|---|
| Standard | 0.90 dollars | 2.70 dollars | 0.045 dollars |
Tencent limited cybersecurity training data to defensive use cases. This is a safety pattern that Asian and Western labs are converging on. The model has undergone specialized training but Tencent has not published a detailed safety report or red team results yet. System cards coming in the coming days will presumably include more detail.
We engineer custom ai engineering and production AI architectures tailored to your existing tech stack.
Tencent plans to release model weights and system cards in the coming days. No exact date has been published. Until then the model is only accessible through Tencent Cloud and third party API providers.
No. Weights are not yet released. Self hosting will be possible once Tencent publishes the weights and system cards. The license terms for the weights have not been announced.
At 0.90 dollars per million input and 2.70 dollars per million output Hy4 is positioned as cost effective. Tencent says it delivers exceptional cost effectiveness despite performance improvements over Hy4 Preview. Hy4 Flash will offer even greater cost efficiency.
Hy4 can interact with command line interfaces and graphical user interfaces. Specialized training on Blender and CAD suggests strength in creative and design software workflows. Tencent has not published specific task examples or success rates yet.
Tencent claims Hy4 ranks first among open weight models on its finance evaluation. No direct head to head benchmark comparison against DeepSeek V3 Qwen3 or other frontier open weights models has been published. Independent verification pending weights release.
Hy4 is the flagship model. Hy4 Flash is scheduled for release soon and will deliver greater cost efficiency for users who prioritize high value applications. Pricing and specifications for Hy4 Flash have not been announced.
WorkBuddy and CodeBuddy will be available for free from November 4 to 15 2026. Hy4 on Tencent Cloud TokenHub is 50 percent off standard pricing through December 31 2026.
Tencent has not published a specific language list. The model is described as delivering stronger overall performance across a diverse set of applications. More detail expected in system cards. Article prepared by Astra Principal AI Systems Engineer Slacode Engineering Journal Data sourced from Tencent official announcement November 4 2026 SiliconANGLE MarkTechPost TechNode AICHINA AISeng Tech and independent summaries. All performance figures are Tencent self reported pending weights release and independent evaluation.
Astra is a content writer at Slacode, creating clear, informative, and engaging content on technology, software development, digital solutions, and business growth. With a focus on accuracy, readability, and practical insights.