Arena.ai
@arena
Hy3 preview (295B A21B) an open source model by @TencentHunyuan is now live on Arena. Evaluate it across Text & Code Arena in Battle mode. Scores incoming soon. https://t.co/5nU6tLSKvK
7retweets88likes
View on XArena.ai Adds Tencent Hy3 Preview for Public Reasoning and Code Benchmarking
Arena· Updated
Arena.ai has added Tencent's Hy3 preview model to its Text and Code Arena leaderboards for public evaluation. This move subjects the 295B-parameter model to blind human testing, providing a verified performance rank against proprietary frontier models.
Arena.ai, a community-driven platform for evaluating AI model performance, added Hy3 preview to its "Battle" mode for text and coding tasks. Developed by Tencent, the model is a massive mixture-of-experts (MoE) architecture (a design activating only a fraction of parameters per request) with 295B total and 21B active parameters.
This benchmarking phase mirrors the recent open-sourcing of Hy3 preview and its integration into hosting platforms like OpenRouter. While self-reported benchmarks claimed high efficiency, Arena's blind testing provides the first human-verified data on how it compares to frontier models in real-world usage.
You can now test the model's reasoning capabilities for free by participating in side-by-side battles on the Arena website. The update adds to recent evaluations like DeepSeek V4 Pro's coding rank, determining if this 21B-active parameter model can match proprietary systems for production agent workflows.
Still wondering? A few quick answers below.
Tencent Hy3 preview is a large-scale foundation model designed for advanced reasoning and autonomous agent tasks. It is the latest iteration of the Hunyuan model family, rebuilt to support complex multi-step actions and logical problem-solving. The model is currently available as a preview version for community feedback before the official release of the final version.
The model uses a Mixture-of-Experts architecture, which contains 295 billion total parameters but only activates 21 billion parameters per token during inference. This design allows the model to maintain high performance and reasoning depth while significantly reducing the computational cost and latency typically associated with running extremely large language models on standard hardware.
Yes, Tencent has open-sourced the weights for the Hy3 preview model. This allows developers and researchers to download, run, and evaluate the model on their own infrastructure. The release is intended to encourage community testing and feedback to help the Tencent Hunyuan team improve the model's performance and safety for its official launch.
Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →
