Tencent's Apache-licensed Hy3 takes on GLM-5.2 at half the size — and wins everywhere except coding

2026-07-06 · Source: VentureBeat · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Advanced, medium

Summary

Tencent's Hunyuan team has released Hy3, a 295-billion-parameter Mixture-of-Experts (MoE) model with 21 billion active parameters, under the permissive Apache 2.0 license. This license change removes previous regional restrictions, making it accessible for global enterprise deployments. Hy3 demonstrates strong performance in agentic search (84.2 on BrowseComp, 91.0 on DeepSearchQA), tool orchestration (79.1 on MCP-Atlas), and long-context retrieval (73.4 on AA-LCR), competing with models two to five times its size. While it trails GLM-5.2 in coding benchmarks, Tencent emphasizes Hy3's improved reliability, reporting a hallucination rate reduction from 12.5% to 5.4% and commonsense error rates from 25.4% to 12.7%. Its deployment footprint is significantly smaller, requiring under 300GB for FP8 weights, making it servable on hardware like Nvidia's H20-3e GPUs.

Key takeaway

For AI Architects and ML Engineers evaluating open-weight models for enterprise deployment, Hy3 presents a compelling option. Its Apache 2.0 license removes previous legal hurdles, while its smaller footprint (under 300GB FP8) and improved reliability (5.4% hallucination rate) offer significant operational advantages over larger models like GLM-5.2, particularly for agentic search and tool orchestration. You should consider Hy3 for reliability-sensitive applications where coding performance isn't the sole criterion, optimizing for cost-effective, export-compliant infrastructure.

Key insights

Tencent's Hy3 offers a reliable, deployable open-weight MoE model with a permissive license, balancing performance with resource efficiency.

Principles

• Permissive licenses drive enterprise adoption.
• Production reliability metrics are key.
• Resource-efficient models can compete.

Method

Tencent refined Hy3 by collecting feedback from over 50 internal product teams post-preview, addressing task execution and interaction issues, and scaling its post-training pipeline.

In practice

• Prioritize models with Apache 2.0 licenses.
• Assess models on hallucination and error rates.
• Consider MoE for compute-constrained agent workloads.

Topics

• Tencent Hy3
• Mixture-of-Experts
• Apache 2.0 License
• Large Language Models
• Model Deployment
• Agentic AI
• Hallucination Reduction

Code references

• Tencent-Hunyuan/Hy3-preview

Best for: CTO, VP of Engineering/Data, MLOps Engineer, Machine Learning Engineer, AI Architect, Director of AI/ML

Related on AIssential

• GLM-5.2: Only a Few Months Behind Commercial Models — GLM-5.2 nears commercial model performance, while VibeThinker-3B shows small models can excel in verifiable reasoning via targeted training.
• Tencent uses product rollout, not just benchmarks, to define Hy3 preview — Tencent's Hy3 preview emphasizes real-world product integration and performance over isolated benchmark claims.
• Mistral Small 4 in 8 mins! — Mistral Small 4 unifies reasoning, coding, and multimodal capabilities for enterprise applications under an Apache 2.0 license.
• Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild — Hy-MT2 delivers high-quality, efficient multilingual translation with strong instruction-following across diverse real-world and resource-constrained scenarios.
• Gemma 4 and what makes an open model succeed — Open model success hinges on ease of use and adaptability, not just benchmark scores.
• Insane Open Source AI Model Just Dropped — GLM 5.1 is an open-source model matching proprietary LLMs in software engineering benchmarks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by VentureBeat.