Tencent’s Hy4 Preview Targets Production-Grade AI

#long-context#coding-agents#open-source-modelstencent-hunyuan-hy4-previewtencenthunyuan hy4 previewtokenhubopenrouterkimi k3

💡A 770B model with 1M context, low pricing, and evidence of self-testing code workflows.

⚡ 30-Second TL;DR

What Changed

Hy4 preview grows from Hy3’s 295B total parameters and 21B active parameters to 770B and 49B respectively, with context extended from 256K to 1M tokens.

Why It Matters

Hy4 preview strengthens Tencent’s position in cost-efficient, open-weight-style model access while emphasizing agentic workflows rather than single-turn generation. Its long context, low pricing, and ability to test and revise outputs could make it attractive for enterprise automation and coding agents, although the preview still requires human review of critical assumptions.

What To Do Next

Run a controlled pilot of Hy4 preview through TokenHub or OpenRouter using your longest coding and document workflows, and compare quality, latency, and cost against your current model.

Who should care:Developers & AI Engineers

Key Points

•Hy4 preview grows from Hy3’s 295B total parameters and 21B active parameters to 770B and 49B respectively, with context extended from 256K to 1M tokens.
•It is available in WorkBuddy, CodeBuddy, Yuanbao, and ima, and can be accessed through Tencent Cloud TokenHub and OpenRouter.
•Pricing is set at 6 yuan per million input tokens, 18 yuan per million output tokens, and 0.3 yuan per million cached tokens.
•In Tencent’s blind test of 203 engineering tasks, Hy4 preview scored 2.99/4, slightly ahead of Kimi K3 at 2.94 and GLM 5.3 at 2.92.
•A hands-on evaluation found it could cross-check expense documents, build a Canvas game with automated tests, and revise implementations after identifying bugs.

🧠 Deep Insight

Background and context from public sources — not the original article. 13 sources cited.

🔑 Enhanced Key Takeaways

•Tencent released the Hy4 model under the Apache 2.0 license, enabling unrestricted commercial use and community-driven development.
•The model was trained using a recursive self-improvement loop where Hy4 assisted in optimizing its own training data strategies and evaluation frameworks.
•Tencent claims a 31.8% increase in end-to-end training throughput as a direct result of the model's contribution to its own development pipeline.
•Hy4 supports standard open-source inference engines including vLLM and SGLang, facilitating easier enterprise self-hosting compared to proprietary-only models.
•Tencent provides an FP8 quantized version of the model alongside the full-precision release to lower hardware requirements for production deployment.

📊 Competitor Analysis▸ Show

Feature	Tencent Hy4	Moonshot Kimi K3	Z.ai GLM-5.3
Architecture	770B MoE (49B active)	Proprietary	Proprietary
Context Window	1M tokens	N/A	N/A
Engineering Score	2.99/4	2.94/4	2.92/4
License	Apache 2.0	Closed	Closed
Cache Efficiency	85% lower cost	Baseline	Baseline

🛠️ Technical Deep Dive

Architecture: Mixture-of-Experts (MoE) design with 770B total parameters and 49B active parameters per token.
Quantization: Native support for FP8 precision to optimize memory footprint and inference speed.
Compatibility: Full support for vLLM and SGLang inference frameworks for production-grade deployment.
Training Methodology: Utilized recursive self-improvement loops where the model contributed to its own data strategy and evaluation framework optimization.

🔮 Future ImplicationsAI analysis grounded in cited sources

Open-weight MoE models will become the standard for enterprise productivity.

The combination of Apache 2.0 licensing and high-efficiency MoE architectures lowers the barrier for companies to replace proprietary APIs with self-hosted solutions.

Recursive self-improvement will shorten AI development cycles by over 20%.

Tencent's reported 31.8% throughput gain suggests that using LLMs to optimize their own training pipelines is a scalable path to increasing R&D efficiency.