Tencent Open-Sources 770B-Parameter Hy4
💡Evaluate Tencent’s new open-source 770B model for million-token context workloads.
⚡ 30-Second TL;DR
What Changed
Hy4 is Tencent Hunyuan’s next-generation large language model preview.
Why It Matters
Hy4 could give developers and researchers access to a very large open-source model for experimentation and deployment. Its million-token-plus context capacity may be particularly useful for long-document analysis, codebase understanding, and other context-heavy workloads.
What To Do Next
Download or access the Hy4 preview and benchmark it on a representative long-context task, such as analyzing your largest documentation set or code repository.
Key Points
- •Hy4 is Tencent Hunyuan’s next-generation large language model preview.
- •The model has 770 billion total parameters.
- •Its context window exceeds 1 million tokens and the model is open-sourced.
🧠 Deep Insight
Background and context from public sources — not the original article. 8 sources cited.
🔑 Enhanced Key Takeaways
- •Hy4 utilizes a Mixture-of-Experts (MoE) architecture with 49 billion active parameters out of its 770 billion total parameter count.
- •The model was trained on specialized datasets derived from Tencent's internal domains, specifically game development, software engineering, finance, and security.
- •Internal blind testing indicates Hy4 outperforms competitors like GLM-5.3 and Kimi K3, achieving an average score of 2.99/4.00.
- •The model has demonstrated advanced scientific reasoning capabilities, notably contributing to the resolution of the three-dimensional Blaschke–Lebesgue geometric problem.
- •Tencent has adopted a rapid development cadence for the Hunyuan series, averaging one major version update every two months since February 2026.
📊 Competitor Analysis▸ Show
| Feature | Hy4 Preview | GLM-5.3 | Kimi K3 |
|---|---|---|---|
| Total Parameters | 770B | N/A | N/A |
| Active Parameters | 49B | N/A | N/A |
| Context Window | 1M+ tokens | N/A | N/A |
| Internal Benchmark Score | 2.99/4.00 | 2.92/4.00 | 2.94/4.00 |
🛠️ Technical Deep Dive
- Architecture: Mixture-of-Experts (MoE) design.
- Parameter Efficiency: 49 billion active parameters per forward pass.
- Deployment: Weights available on Hugging Face, GitHub, ModelScope, and Gitcode.
- API Integration: Accessible via Tencent Cloud TokenHub and OpenRouter.
- Pricing Model: 6 yuan per million input tokens; 18 yuan per million output tokens.
🔮 Future ImplicationsAI analysis grounded in cited sources
⏳ Timeline
📎 Sources (8)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: Pandaily ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
Weekly AI briefing
One email a week. Unsubscribe anytime.