Hy4 Review: Tencent Vision and Computer-Use Model Pricing, Capabilities and Platform Access

Hy4 review examines Tencent's full release of its 770B-parameter MoE model on November 4, 2026. The model supports 1M-token context, vision processing for images and screenshots, and computer-use actions across CLI and GUI environments. It builds directly on the August 28 text-only preview with added multimodal and agent capabilities.

Hy4 Overview and Core Capabilities

Tencent developed Hy4 for productivity tasks in coding, finance, office work, research, and game development. The architecture uses 49B active parameters per token in its preview configuration, with a maximum output of roughly 64K to 65K tokens. Vision functions handle fine-grained chart analysis and mathematical visual reasoning, while computer-use training covers tools such as Blender and CAD alongside multi-application workflows.

Key features include tool calling, JSON output mode, and configurable reasoning levels set to none, low, medium, or high. The model integrates directly into Tencent products including WorkBuddy, CodeBuddy, Yuanbao, ima, and WeCom. It also appears on third-party platforms such as OpenRouter under the identifier tencent/hy4 and through Tencent Cloud TokenHub with an OpenAI-compatible API.

Benchmarks reported by Tencent place Hy4 at 75.2 on APEX-Agents v1.1 for office and agent tasks, ahead of Claude Opus 5.5 in the company's internal runs. On OSWorld 2.1 computer-use evaluation it scores 78.9, compared with 81.8 for Claude and 79.6 for Kimi K3.1 Pro. Vision performance reaches 78.6 on MMMU-Pro, while Terminal-Bench 4.0 shows 44.4. Internal finance evaluations with 213 professionals across 579 tasks yield an average of 3.04 out of 4.

Pricing, Access Options, and Performance Comparison

Pricing on Tencent Cloud TokenHub lists at approximately $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached tokens. A 50 percent launch discount remains available through December 31, 2026, halving those figures. Third-party platforms such as AI/ML API quote slightly higher rates near $0.88 input and $2.65 output. Reasoning tokens count toward output billing on all supported platforms.

Access requires an account on Tencent Cloud TokenHub or OpenRouter. Limited free quotas appeared in select Tencent applications from November 4 through November 15. Model weights and the full system card have not yet been released, though the earlier preview carried an Apache 2.0 license. Open-weight release and a lower-cost Hy4 Flash variant are expected in the coming weeks.

Security details remain limited to standard Tencent Cloud compliance statements. No independent audits or known incidents have surfaced in the first 24 hours after launch. Hallucination rates improved internally from 15.76 percent in the preview to 5.45 percent in the full release.

Independent benchmarks and real-world agent performance data are still pending as of November 5, 2026. User reviews and complaints have not accumulated in volume yet due to the recent release. Early developer feedback on OpenRouter and AI/ML API forums notes strong gains on long-horizon agent tasks but flags occasional grounding errors in complex multi-app sequences.

Limitations include the absence of released weights, lower scores on Terminal-Bench 4.0 relative to leading models, and reliance on vendor-reported metrics for most evaluations. No public information exists on data retention policies or third-party model training use at this stage. Potential issues center on the need for further verification of computer-use reliability outside Tencent's internal test environments.