Tencent Hy3 Release Notes

Tencent has released Hy3, an open-source model designed to outperform similar-sized models and compete with flagship open-source models that possess 2-5x more parameters. The model is released under the Apache 2.0 license and is available on GitHub, HuggingFace, ModelScope, and AtomGit.

Enhanced Agent and Reasoning Capabilities

Hy3 delivers significant gains in reasoning, agentic, and long-context tasks. In blind evaluations conducted by 270 experts using real-world work tasks, Hy3 scored 2.67/4, surpassing GLM-5.1's score of 2.51/4. The model's performance advantages are most pronounced in frontend development, data and storage, and CI/CD tasks.

Production-Grade Reliability and Hallucination Reduction

Hy3 introduces several architectural and training improvements to increase reliability for production environments:

• Tool Call Stability: The model has been brought to production-grade standards for tool configurations and output constraints, improving error recovery and efficiency. On SWE-Bench Verified, accuracy variance across different agent scaffoldings (such as CodeBuddy, Cline, and KiloCode) remains within 4%.
• Reduced Hallucinations: Through fine-grained data cleaning and training constraints, Hy3's hallucination rate dropped from 12.5% to 5.4%, while commonsense error rates decreased from 25.4% to 12.7%.
• Context Retention: Joint optimization of SFT and RL has improved coreference resolution, ellipsis recovery, and multi-turn constraint inheritance. Internal multi-turn test issue rates dropped from 17.4% to 7.9%.

Operational Efficiency and Token Usage

Internal testing at WorkBuddy demonstrated a task success rate increase from 72% (Hy3 preview) to 90% with Hy3, alongside a 34% reduction in average completion time. Hy3 also exhibits superior token efficiency compared to GLM-5.2, using 47.4% fewer tokens for document processing and 49% fewer for presentation creation.

API Pricing

Hy3 is offered with the following pricing per 1M tokens:

Input	Output	Cached Input
1 RMB	4 RMB	0.25 RMB

Community Perspectives and Benchmarks

Community feedback on Hy3 is mixed, with some users noting its efficiency and cost-effectiveness while others question its real-world utility compared to competitors.

"This model is shockingly small for how capable it is... I wouldn't be surprised if this becomes a popular local model."

Some users have compared Hy3 to other models like DeepSeek V4 Flash, noting that it is slightly larger but potentially more capable on certain benchmarks. Others have expressed skepticism regarding its performance in practical applications, suggesting it may be "benchmaxxed" or underperform compared to dense Gemma models.

One user reported that Hy3's performance is close to Sonnet 5 and better than GPT-5.4-mini for their specific use cases, emphasizing its low cost as a primary driver for adoption.