SyncFlo AI Logo
← Back to News Feed
OPEN-WEIGHT FRONTIER AI • TENCENT • HY4 • 770B MOE • COMPUTER USE • 1M CONTEXT • MULTIMODAL LATENT • WORKBUDDY • TOKENHUB

Tencent Launches Hy4: 770B Mixture-of-Experts Open-Weight Frontier Model with Native Multimodal Latents and OS-Level Computer-Use Agents

By SyncFlo AI Editorial Team · · 13 min read
Cinematic visualization of Tencent Hy4 770B Mixture-of-Experts neural architecture with glowing amber expert routing cores, native multimodal vision streams, and computer-use agent navigation cursors.
Tencent unveils the full release of Hy4, an open-weight 770B parameter MoE frontier model activating 49B parameters, featuring native vision integration and autonomous OS-level computer-use execution. | Source: Tencent AI Lab & Tencent Cloud. Visual: SyncFlo AI News

SHENZHEN, China — November 11, 2026 — In a monumental demonstration of open-weight artificial intelligence compute scaling, global technology conglomerate Tencent Holdings Ltd. has officially released the full production version of Hy4, its flagship 770-billion-parameter Mixture-of-Experts (MoE) foundation model. Activating approximately 49 billion parameters per inference token, Hy4 represents a massive 2.6x scale expansion over its predecessor while dramatically reducing operational token latency through cutting-edge sparse-expert routing algorithms.

Following an initial text-focused preview release earlier this year, the full November launch of Hy4 introduces two transformative technological capabilities to the open-weight ecosystem: native multimodal visual understanding baked directly into its primary latent space, and an autonomous "Computer-Use" agent framework engineered to operate graphical operating systems, terminal environments, and professional creative applications like Blender and Autodesk CAD with human-grade spatial precision.

1. Architectural Blueprint: 770B MoE with 1M Token Context

Rather than relying on brute-force dense parameter activations that make frontier deployment prohibitively expensive for private enterprise data centers, Tencent designed Hy4 as an ultra-sparse Mixture-of-Experts system:

  • Sparse Activation Ratio: Out of 770 billion total model parameters, exactly 49 billion active parameters are routed per forward token pass across 128 top-k routing experts, ensuring that latency matches dense 50B models while cognitive capacity scales to frontier heights.
  • 1,000,000-Token Native Context Window: Hy4 effortlessly ingests entire enterprise codebases, full multi-year financial audit logs, and hours of structured video telemetry in a single session, scoring 99.4% retrieval accuracy across 1M tokens.
  • Native Latent Multimodality: In contrast to standard Vision-Language Models (VLMs) that bridge a pre-trained frozen vision encoder (like CLIP or SigLIP) with a separate LLM through projection layers, Hy4 processes visual tokens natively within the core transformer backbone. This unified latent representation eliminates translation degradation when analyzing complex blueprints, scientific graphs, and user interfaces.

Tencent Hy4 Technical Benchmarks & System Metrics

Cognitive & Agentic Benchmarks
  • OSWorld (Autonomous Computer-Use): Achieved 44.8% success across complex multi-application desktop workflows, leading all open-weight systems.
  • SWE-bench Verified: 54.2% resolution rate across automated software repository bug fixing and test-suite generation.
  • MMMU (Multimodal Understanding): 69.8% on multidisciplinary reasoning across technical schematics and medical imagery.
  • Math-500: 93.6% accuracy on complex graduate competition mathematics using test-time step-level verification.
Deployment, Silicon & API Economics
  • TokenHub Cloud Pricing: $0.65 per million input tokens / $2.40 per million output tokens, delivering aggressive price-performance.
  • Open-Weight Rollout: Base weights released under Apache 2.0 license with staged releases for multimodal and agent checkpoints.
  • Hardware Footprint: 4-bit and 8-bit quantized checkpoints fit within single 8x H200/B200 node configurations for sovereign on-prem hosting.
  • Ecosystem Integrations: Deep integration with Tencent WorkBuddy, CodeBuddy, and WeChat Work Dayuan assistant.
"Open-weight intelligence must not remain confined to text synthesis. With Hy4, we are delivering an open, sovereign foundation capable of directly seeing the digital workspace and autonomously operating the software tools that humans rely upon daily—from visual software engineering to high-precision industrial CAD."
— Senior Leadership, Tencent AI Lab & Tencent Cloud Engineering

2. Autonomous Computer-Use: Bridging Code and GUI

A standout capability of the Hy4 release is its native support for autonomous computer control. By treating the computer screen as an interactive sensory canvas, Hy4 does not merely predict API calls; it inspects visual pixel buffers, computes exact coordinate clicks, issues keyboard commands, scrolls through documentation, and handles unexpected OS modal dialogues.

In live developer demonstrations, Hy4 was tasked with opening Blender, importing a rough 3D scan of an automotive component, modifying the geometric mesh to reduce aerodynamic drag, rendering the output file, and pushing the final assets to an enterprise Git repository—all without human keystrokes. This native dexterity makes Hy4 an indispensable tool for automating legacy enterprise software that lacks modern REST APIs.

3. Enterprise Ecosystem: WorkBuddy, CodeBuddy & Dayuan

Tencent is pairing the open-weights release with comprehensive commercial distribution across its enterprise software umbrella:

  • Tencent CodeBuddy: Supercharged by Hy4's 1M-token context and 54.2% SWE-bench score, CodeBuddy acts as an autonomous pair programmer capable of multi-file refactoring and automated security vulnerability patching.
  • WeChat Work Dayuan: Over one billion enterprise workers on WeChat Work gain access to Dayuan, an AI agent capable of scheduling cross-departmental operations, generating financial reports, and approving operational workflows.
  • Tencent Cloud TokenHub: Enterprise developers can instantly consume Hy4 via high-availability cloud APIs, with support on third-party aggregators including OpenRouter.

4. Frontier Alignment, Biosecurity & Responsible Deployment

With open-weight models approaching frontier capabilities, safety and governance have taken center stage. Tencent's chief AI safety scientists highlighted extensive evaluations conducted to prevent malicious misuse in biological synthesis, kinetic operations, and unauthorized network exploitation.

To ensure safe proliferation, Tencent adopted a staged rollout: while base reasoning weights are accessible under Apache 2.0, high-risk agentic action modules are governed by developer registration and cryptographic license auditing, preventing autonomous weaponization while maintaining full open-weight inspectability for scientific researchers.

5. SyncFlo AI Analysis: The Open-Weight Sovereignty Revolution

At SyncFlo AI, we view Tencent Hy4 as a watershed moment for enterprise data sovereignty. For organizations operating under stringent regulatory jurisdictions—such as sovereign wealth funds, healthcare systems, and defense contractors—closed proprietary API models pose persistent compliance and vendor lock-in challenges.

The arrival of an open-weight 770B MoE model with native Computer-Use allows enterprises to host frontier-class intelligence completely within their on-premises or sovereign cloud infrastructure. Through SyncFlo AI's enterprise agent orchestration framework, businesses can fine-tune Hy4 on proprietary operational data, deploying sovereign agent workforces that maintain absolute privacy while outperforming legacy closed-source benchmarks.