Particle.news

Tencent Releases Hy4, an Open-Source Multimodal Model That Can Operate Software

The company says it will publish model weights and system documentation soon to allow outside audit and broader use on cloud and third-party APIs.

Overview

  • Tencent announced Hy4 as the next Hy-family model with new visual reasoning and computer-control abilities that let it interpret complex images, read charts, solve visual math, and interact with command-line and graphical tools.
  • The company made Hy4 available across its cloud and collaboration products and via APIs that support third-party services, and it says model weights and system cards will be published in the coming days.
  • Tencent reported internal evaluation gains including a fall in hallucination rate to 7.2 percent from 12.95 percent and a drop in unsupported-assumption errors to 5.6 percent alongside high task pass rates and benchmark scores.
  • Tencent says it trained Hy4 on defensive cybersecurity material, excluded offensive exploitation content from training, and added safeguards for biological and cyber risks, but those measures and safety claims come from company testing.
  • Pricing remains at about $0.90 per million input tokens and $2.70 per million output tokens, Tencent plans a smaller lower-cost Hy4 Flash variant, and the model’s ability to automate finance, engineering, and development workflows raises both productivity opportunities and safety concerns pending external review.