tencent/Hy3

2026-07-06 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, quick

Summary

Tencent has released Hy3, an Apache 2.0 licensed 295B-parameter Mixture-of-Experts (MoE) model developed by the Tencent Hy Team. This model features 21B active parameters and 3.8B MTP layer parameters. Following a late April preview and post-training with higher quality data, Hy3 now outperforms other models of similar size and competes effectively with flagship open-source models that possess 2-5 times more parameters. It also demonstrates substantial utility improvements across various products and productivity tasks. The full-sized model occupies 598GB, while an FP8 quantized version is available at 300GB. Hy3 boasts an extensive 256K context length and is currently offered for free on OpenRouter until July 21st.

Key takeaway

For Machine Learning Engineers evaluating large language models, Hy3 presents a compelling option. Its 295B-parameter MoE architecture, with only 21B active parameters, rivals models 2-5x larger in performance. You should consider testing Hy3 for productivity tasks, especially given its 256K context length and current free access on OpenRouter until July 21st. This could significantly reduce inference costs while maintaining high utility.

Key insights

Tencent's Hy3 MoE model delivers competitive performance against much larger models while maintaining a more efficient active parameter count.

Principles

• MoE architectures can achieve high performance efficiently.
• Post-training with high-quality data improves model utility.

In practice

• Utilize Hy3 for diverse productivity tasks.
• Access Hy3 free on OpenRouter until July 21st.
• Employ FP8 quantization for reduced model size.

Topics

• Mixture-of-Experts
• Large Language Models
• FP8 Quantization
• Tencent Hy3
• OpenRouter
• Apache 2.0 License

Best for: NLP Engineer, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• 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.
• Why o3 Gets Smarter the Longer You Let It Think: Test-time compute explained from first principles. — Test-time compute scaling dramatically improves AI model performance without retraining, challenging traditional scaling assumptions.
• Alibaba Qwen Team Releases Qwen3.5-397B MoE Model with 17B Active Parameters and 1M Token Context for AI agents — Qwen3.5 MoE model achieves 400B-class reasoning with 17B-level inference speed and a 1M token context.
• 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.
• DeepSeek-V3: A Large-Scale MoE Pretraining Benchmark for MLPerf Training v6.0 — MLPerf Training v6.0 now benchmarks MoE LLMs, reflecting industry shifts to sparse computation and specialized architectures.
• I Ran the 3B Model That Beat Gemini 3 Pro at Olympiad Math — It Shouldn't Work — A small, cost-effective 3B model achieved top-tier verifiable reasoning performance, challenging assumptions about model size and capability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.