Tencent releases Hy3, upgrading its Hunyuan model from preview to general availability

Tencent has released Hy3, the general-availability version of its Hunyuan large language model, following up on the Hy3 Preview it launched on OpenRouter and Hugging Face in April. The model card describes it as a scaled-up production release built on feedback from over 50 product teams that used the preview.

What's new

Hy3 is a 295-billion-parameter Mixture-of-Experts model with 21 billion active parameters, using a differentiated expert-size design — 192 experts with the top 8 activated per token — where tokens of varying difficulty are routed to experts with different capacities. Tencent says this approach, tested under matched compute budgets, "significantly outperforms traditional MoE." The model supports a 256K token context window and BF16 precision, and is released under the Apache License 2.0.

Tencent's model card states the GA version "has made remarkable progress and can now serve as a reliable, cost-effective model option," following the improvements made after the April preview: "Following the Hy3 Preview launch in late April, we gathered feedback from 50+ product teams. We fixed various issues in task execution and interaction, and improved both the quality and scale of our post-training pipeline." On benchmarks, Tencent cites a blind expert evaluation where Hy3 scored 2.67 out of 4, ahead of GLM-5.1's 2.51, and describes overall performance as rivaling flagship open-source models with two to five times as many parameters, with particular gains in complex reasoning, instruction following, coding, and agent tasks.

Context

The Hy3 Preview briefly topped OpenRouter's daily API usage rankings after its April debut, as developers took advantage of a limited-time free access window. That preview was the first model released after Tencent tore down and rebuilt its Hunyuan architecture from scratch, a process the company has said took less than three months from infrastructure reconstruction to release. The GA release lands amid intense competition among Chinese open-weight labs — DeepSeek, Alibaba's Qwen, Moonshot's Kimi, and Z.ai's GLM among them — all shipping large MoE models that trade raw parameter count for cost-efficient performance close to proprietary frontier systems.

Why it matters

Moving from a time-limited, free preview to a fully licensed Apache 2.0 release, refined with direct input from more than 50 product teams, signals Tencent is treating Hy3 as core infrastructure rather than a one-off research drop. The differentiated expert-size MoE design is a notable architectural choice distinct from the uniform-expert MoE approach most other labs use, and if the efficiency gains Tencent claims hold up under independent testing, it adds another data point to the broader trend of open-weight labs closing the gap with closed frontier models on agent and reasoning tasks at a fraction of the deployed parameter count.