Tencent Hy4 Bets on Open-Source Pragmatism
💡A 770B open model with 1M context targets real production workflows—but its reasoning trade-offs matter.
⚡ 30-Second TL;DR
What Changed
Hy4 preview has 770B parameters, a 1M-token context window, and open-weight positioning.
Why It Matters
Hy4 strengthens the open-model competition by combining a very large parameter count, long context, and low pricing with Tencent’s application ecosystem. However, its reliance on community-originated architectural ideas and weaknesses in reasoning consistency may limit its standing as a foundational model innovator.
What To Do Next
Run Hy4 preview against your existing agent and coding evaluation suite, specifically measuring tool-call success, 1M-context retrieval, latency, and repeated self-verification.
Key Points
- •Hy4 preview has 770B parameters, a 1M-token context window, and open-weight positioning.
- •The architecture uses DeepSeek’s DSA and MTP, Zhipu’s IndexCache, and a lighter iHC residual design inspired by mHC.
- •Benchmark gains are concentrated in agentic, tool-use, and engineering tasks rather than pure mathematics or long-horizon reasoning.
- •Tencent positions Hy4 as a productivity model co-designed for WorkBuddy and CodeBuddy.
- •The model reportedly suffers from excessive self-verification and slow reasoning on complex tasks.
🧠 Deep Insight
Background and context from public sources — not the original article. 12 sources cited.
🔑 Enhanced Key Takeaways
- •Hy4 utilizes a Mixture-of-Experts (MoE) architecture with 770B total parameters and 49B active parameters per token.
- •The model was developed using a recursive self-improvement loop where Hy4 optimized its own training strategies and operators.
- •Tencent reported a 31.8% improvement in end-to-end inference throughput by using the model to autonomously identify and resolve system bottlenecks.
- •Training data was curated specifically from Tencent's internal domains, including gaming, finance, and security, rather than relying on generic web-scale datasets.
- •The model is released under the Apache 2.0 license, with weights available on Hugging Face, ModelScope, GitCode, and CNB.
📊 Competitor Analysis▸ Show
| Feature | Hy4 Preview | GLM-5.3 | Kimi K3 |
|---|---|---|---|
| Internal Expert Score | 2.99 | 2.92 | 2.94 |
| Active Parameters | 49B | N/A | N/A |
| Pricing (Input/Output) | 6/18 CNY per M tokens | N/A | N/A |
🛠️ Technical Deep Dive
- Architecture: Mixture-of-Experts (MoE) with 770B total parameters and 49B active parameters.
- Inference: Achieved 31.8% throughput gain via autonomous system bottleneck optimization.
- Quantization: Official release includes FP8 quantized versions.
- Training Methodology: Recursive self-improvement loop for training data and operator optimization.
- Context Window: 1M tokens supported via IndexCache and iHC residual design.
🔮 Future ImplicationsAI analysis grounded in cited sources
⏳ Timeline
📎 Sources (12)
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: 虎嗅 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
Weekly AI briefing
One email a week. Unsubscribe anytime.