Tencent Hy3 Release Notes

Tencent has released Hy3, an open-source model designed to rival flagship models with 2-5x more parameters. The model focuses on enhancing agentic capabilities, reducing hallucinations, and improving token efficiency for productivity tasks such as coding and financial modeling.

Enhanced Agent and Reasoning Capabilities

Hy3 delivers significant gains in reasoning, agentic, and long-context tasks by scaling up reinforcement learning (RL) and improving the quality and diversity of post-training data. In a blind evaluation involving 270 experts, Hy3 scored 2.67/4, outperforming GLM-5.1 (2.51/4). The most substantial performance gains were observed in frontend development, data and storage, and CI/CD tasks.

Production-Grade Reliability and Hallucination Reduction

Hy3 introduces several architectural and training improvements to make the model more reliable for production environments:

Tool Call Stability: The model now meets production-grade standards for tool configurations and output constraints, with improved error recovery. On SWE-Bench Verified, accuracy variance across different scaffoldings (such as CodeBuddy, Cline, and KiloCode) remains within 4%.
Factuality: Through fine-grained data cleaning and training constraints, Hy3 reduced its hallucination rate from 12.5% to 5.4% and lowered commonsense error rates from 25.4% to 12.7%.
Context Retention: Joint optimization of SFT and RL improved coreference resolution and multi-turn constraint inheritance. Internal multi-turn test issue rates dropped from 17.4% to 7.9%, and the model showed marked improvement on long-dialogue evaluations like MRCR.

Productivity Gains and Token Efficiency

In internal testing with WorkBuddy, Hy3 increased the task success rate from 72% (Hy3 preview) to 90%. Average completion time decreased by 34%.

Hy3 also demonstrates superior token efficiency compared to GLM-5.2. In common tasks, Hy3 used 47.4% fewer tokens for document processing and 49% fewer tokens for presentation creation.

Availability and Pricing

Hy3 is open-sourced under the Apache 2.0 license and is available on GitHub, HuggingFace, ModelScope, and AtomGit.

API Pricing (per 1M tokens):

Input	Output	Cached Input
1 RMB	4 RMB	0.25 RMB

Community Feedback and Perspectives

Community reactions to the Hy3 release are mixed, with some users praising its efficiency and others questioning its real-world performance relative to benchmarks.

Performance and Value

Some users noted that the model is "shockingly small for how capable it is," comparing it favorably to DeepSeek-V4 Flash and noting it may become a popular local model. Others found it to be a strong, cost-effective alternative for those needing a FOSS model, with one user stating it is "close enough to sonnet 5 in performance that I didn't notice much of a gap."

Critical Counterpoints

Conversely, some users expressed skepticism regarding the benchmark results and the actual utility of the model:

"I feel like I'm taking crazy pills with hy3, it's either benchmaxxed to hell and back or skill issue on my part but I'd rather use dense gemma."

Other critics pointed out potential issues with the model's documentation and user interface, noting that the demo page is "janky" and the front page fails to explain the product's value proposition clearly.

Comparative Analysis

Users comparing Hy3 to other models noted varying results. While some appreciated the speed and instruction-following capabilities, others argued it is "worse than GLM-5.2" and "much more expensive than deepseek v4," suggesting a tough bargain for some use cases.