🔥36氪•Stalecollected in 14h
Tencent AI accelerates with top-ranked Hy3 preview
💡Tencent's Hy3 is currently dominating OpenRouter usage; see if it fits your production agent stack.
⚡ 30-Second TL;DR
What Changed
Hy3 preview leads OpenRouter in token usage for three consecutive weeks.
Why It Matters
Tencent's aggressive AI agent deployment and model optimization indicate a shift toward high-frequency, specialized AI application delivery in the enterprise market.
What To Do Next
Benchmark your current agent workflows against the Hy3 preview API to evaluate its performance in coding and context-heavy tasks.
Who should care:Developers & AI Engineers
Key Points
- •Hy3 preview leads OpenRouter in token usage for three consecutive weeks.
- •Tencent has launched dozens of general and vertical AI agents in 2026.
- •Hunyuan model underwent a full rebuild in under three months to improve coding and context capabilities.
🧠 Deep Insight
Web-grounded analysis with 13 cited sources.
🔑 Enhanced Key Takeaways
- •Tencent's Hy3 preview is a Mixture-of-Experts (MoE) model, featuring 295 billion total parameters with only 21 billion activated per token, and supports an extensive 256K token context window.
- •The recent rebuild of the Hunyuan model's pre-training and reinforcement learning infrastructure, overseen by former OpenAI researcher Yao Shunyu, was specifically geared towards enhancing practical, real-world applications and business needs, moving beyond a sole focus on benchmark scores.
- •Tencent is developing a 'top-secret' AI agent for its WeChat platform, aiming to deeply integrate with its vast mini-program ecosystem to automate complex, multi-step tasks like ride-hailing and food delivery for its 1.4 billion users.
- •The Hy3 preview model demonstrates a 40% improvement in inference efficiency and has proven capable of reliably powering complex agent workflows involving up to 495 steps across various use cases, including document processing and data analysis.
- •Tencent offers the Hy3 preview through its Tencent Cloud TokenHub with competitive pricing, starting at approximately USD 0.18 per million input tokens, indicating a strong push for broad commercial adoption.
📊 Competitor Analysis▸ Show
Competitor Analysis: Tencent Hy3 Preview
| Feature/Metric | Tencent Hy3 Preview | Claude Opus 4.6 | GPT-5.4 | GLM-5 | Kimi-K2 | Llama 3.1 405B (Hunyuan-Large comparison) |
|---|---|---|---|---|---|---|
| Architecture | MoE (295B total, 21B active) | Closed-source, likely MoE | Closed-source, likely MoE | MoE | MoE | MoE (389B total, 52B active) |
| Context Window | 256K tokens | Large | Large | Large | Large | 256K tokens |
| Inference Efficiency | 40% improvement over previous versions | - | - | - | - | Lower processing requirement than dense models of similar size |
| Pricing (Input) | ~$0.18/million tokens (RMB 1.2) | - | - | - | - | Free for developers outside EU with <100M monthly users (Hunyuan-Large) |
| SWE-bench Verified | 74.4% | 80.8% | 78.6% | Competitive | - | - |
| Terminal-Bench 2.0 | 54.4% | - | - | - | - | - |
| WideSearch | 70.2% | - | - | Competitive | Close to Kimi-K2 | - |
| MMLU | - | - | - | - | - | 88.4% (Hunyuan-Large) vs 85.2% (Llama 3.1 405B) |
| AlpacaEval 2 | - | - | - | - | - | 51.8% (Hunyuan-Large-Instruct) vs 50.5% (DeepSeek 2.5 Chat) |
| Agent Capabilities | Powers complex workflows up to 495 steps | - | - | - | - | State-of-the-art agentic performance (Hunyuan-A13B) |
| Key Focus | Production-grade agent workloads, complex reasoning, coding, context learning | Advanced reasoning | Advanced reasoning | General-purpose LLM | General-purpose LLM | General-purpose, open-source, efficiency |
🛠️ Technical Deep Dive
- Model Architecture: Mixture-of-Experts (MoE) model.
- Parameter Count: 295 billion total parameters, with 21 billion parameters activated per token during inference.
- Context Window: Supports a native 256,000 token context window.
- Reasoning Capabilities: Integrates both 'fast thinking' and 'slow thinking' capabilities for enhanced complex reasoning.
- Speculative Decoding: Utilizes a 3.8 billion parameter MTP (Multi-Path Transformer) layer for speculative decoding, aiming for lower latency and smaller batch sizes.
- Internal Structure: Comprises 80 transformer layers, with 192 routed experts (top-8 selection) and 1 shared expert.
- Attention Mechanism: Employs Grouped Query Attention (GQA) with 64 heads over 8 Key-Value (KV) heads.
- Inference Modes: Offers three dynamic inference modes:
no_think,think_low, andthink_high, allowing for a trade-off between latency and reasoning depth. - Infrastructure Rebuild: The model emerged from a rebuilt pre-training and reinforcement learning infrastructure, emphasizing stability and practical business needs.
- Agent Framework Integration: Supports integration with popular open-source agent frameworks such as OpenClaw, OpenCode, and KiloCode.
- Deployment Requirements: To serve on 8 GPUs, requires high-memory GPUs like NVIDIA H200, AMD MI300X/MI325X (192 GB), or AMD MI350X/MI355X (288 GB).
🔮 Future ImplicationsAI analysis grounded in cited sources
Tencent's deep integration of AI agents into WeChat will transform the platform into a dominant intelligent task center in China.
By leveraging WeChat's 1.4 billion users and extensive mini-program ecosystem, the AI agent can automate a wide array of daily tasks, shifting competition from conversational AI to practical execution.
Tencent's focus on MoE architecture and inference efficiency with Hy3 preview will drive broader commercial adoption of its AI models.
The model's 40% improvement in inference efficiency and competitive pricing on Tencent Cloud's TokenHub make it a cost-effective and practical solution for diverse real-world business applications.
Tencent's strategy, guided by former OpenAI researcher Yao Shunyu, signals a shift in the AI industry towards practical, production-ready models over pure benchmark chasing.
The rebuild of the Hunyuan model's infrastructure and the design of Hy3 preview prioritize real-world usability and business needs, suggesting a more pragmatic approach to AI development.
⏳ Timeline
2023-09
Tencent debuts Hunyuan foundation model with over 100 billion parameters and 2 trillion tokens, integrated into 50 Tencent products.
2024-11
Tencent releases Hunyuan-Large, an open-source Mixture-of-Experts model with 389 billion total parameters and 52 billion active parameters.
2024-12
Tencent open-sources HunyuanVideo, a 13 billion parameter video generative model.
2025-09
Tencent open-sources Hunyuan Image 3.0, an 80B total parameter (13B active) text-to-image MoE model.
2026-03
Tencent's WeChat team begins developing a high-priority AI agent product for integration with mini-programs, with phased testing scheduled for mid-2026.
2026-04
Tencent launches and open-sources the Hy3 preview model, following a rebuild of its pre-training and reinforcement learning infrastructure.
📎 Sources (13)
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: 36氪 ↗