88Z AI, 6 July — Tencent Hunyuan has officially released the Hy3 large model. By improving training data quality and scaling reinforcement learning compute, it achieves significant progress in reasoning, agents, and long-context tasks. Its performance rivals flagship models with 2-5 times more parameters, while substantially reducing hallucination rates, common sense errors, and multi-turn issue rates.
At a glance
- Hy3 performance rivals flagship models with 2-5x more parameters
- Hallucination rate drops to 5.4%, common sense errors to 12.7%
The update
Tencent Hunyuan has officially launched the Hy3 large model. Building on the preview version, it focuses on enhancing the quality of post-training data and expanding the scale of reinforcement learning compute. This improvement significantly boosts Hy3's performance in key tasks such as logical reasoning, agent interaction, and long-context processing, allowing it to rival flagship models with 2 to 5 times more parameters. In terms of accuracy, Hy3 has achieved optimization in several key metrics. The model's hallucination rate dropped sharply from 12.5% to 5.4%, common sense errors decreased from 25.4% to 12.7%, and the issue rate in multi-turn dialogues fell from 17.4% to 7.9%. These figures indicate that Hy3 has significantly improved output reliability and stability while maintaining high efficiency.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 腾讯混元Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.