2026-11-16 15:20
Introduction: Only 12B activated out of 224B total parameters, with 1M context support exceeding 150 tps. What is Hy4 flash? Hy4 flash is Tencent Hunyuan's ultra-high cost-performance large model.
What is Hy4 flash?
Hy4 flash is an ultra-high cost-performance large model launched by Tencent Hunyuan, featuring 224B total parameters with only 12B activated, supporting multimodal processing and up to 1M context length. Built on hybrid attention, LatentMoE, and Engram architectures, it drastically reduces KV cache size and achieves decoding speeds exceeding 150 tps. Trained via self-distillation and reinforcement learning, it reduces output length by 50% and delivers performance slightly superior to Hy4 preview. Already deployed on TokenHub, Yuanbao, WorkBuddy, and other platforms, it is ideal for daily tasks such as document processing, PPT generation, and deep search.

Key Features of Hy4 flash
- Deep Search: Enables low-cost multi-round retrieval and information integration, delivering source-tracked answers quickly.
- Document Processing: Supports comprehension, editing, summarization, and format repair of long documents (up to 1M context).
- PPT Generation & Editing: Directly generates and iteratively refines presentations based on user requirements, supporting human-machine collaborative writing.
- Table Repair & Data Analysis: Automatically detects and fixes table issues, completing data organization tasks.
- Simple Frontend Rendering: Generates lightweight web pages and visual interfaces to meet everyday presentation needs.
- Computer Use: Performs multi-step GUI tasks such as system configuration, software download and installation using real client environments.
- Visual Understanding & Audio/Video Processing: Native multimodal support enables parsing of images, audio, and video content.
- Agent Collaboration: Independently plans tool invocation for multi-step execution in scenarios like enterprise information retrieval and root cause analysis of bugs.
Technical Principles of Hy4 flash
- Sparse Mixture-of-Experts Architecture (LatentMoE + 12B Activation): While the total model has 224B parameters, only 12B are activated per inference. Latent space routing concentrates computation on the most relevant experts, maintaining high capability while compressing single-inference compute costs to levels comparable to small models.
- Hybrid Attention & Engram Memory Mechanism: Incorporates hybrid attention layers and a 64B Engram parameter block as external memory, replacing traditional long KV caches. This reduces KV cache volume to just 1/24 of Hy3 and 1/6 of Hy4, making 1M-length context inference feasible within manageable GPU memory constraints.
- Self-Distillation & Length-Optimized Reinforcement Learning: Uses the flagship Hy4 model as a teacher for self-distillation, achieving near-Hy4-level intelligence. Reinforcement learning specifically optimizes end-to-end tasks, reducing output length by 50%, directly cutting token consumption and latency.
- Workflow-Oriented Evaluation & Training Loop: Employs a self-built WorkIndex benchmark set (8 internal task categories + human evaluation) based on real WorkBuddy workflows, focusing on outcome quality, single-turn output length, and end-to-end execution time to ensure optimizations align with daily office use cases.
- Extreme Inference Optimization (megakernel + 4bit Quantization + DFly): On the inference side, megakernel operator fusion minimizes kernel launch overhead; weights use 4bit quantization to reduce GPU memory and bandwidth pressure; combined with DFly speculative/parallel decoding techniques, it achieves decoding speeds over 150 tps, enabling cost-effective deployment at scale.
How to Use Hy4 flash
- Method One: Use Directly in Products (Zero Barrier)
- Open any integrated platform such as Yuanbao App / Web Version, WorkBuddy, ima, Enterprise WeChat, QQ, or Marvis.
- Select or confirm that Hy4 flash is currently active in the model selection menu.
- Submit natural language tasks directly, e.g., “Generate a PPT for me” or “Summarize this document,” and wait for results.
- Method Two: API Integration (For Developers)
- Log into Tencent Cloud Console, activate TokenHub service, and create an API Key.
- Select the Hy4 flash model endpoint and verify the discounted rate (50% off until end of November): input cost at 0.5 RMB/million tokens, output at 2 RMB/million tokens.
- Use OpenAI-compatible SDKs to send requests, setting base_url to the TokenHub endpoint and model to Hy4 flash.
- Pass messages and multimodal content such as files or images; receive streaming responses.
- Monitor token usage in the console and decide whether to switch to Hy4 for complex, long-running tasks.
Core Advantages of Hy4 flash
- Ultra-Fast Inference: Decoding speed exceeds 150 tps; on WorkBuddy, task duration is reduced by 70% compared to Hy4, making it the fastest in blind testing for task completion time.
- Extremely Long Context: Supports up to 1M tokens context, yet GPU memory cost for long-document processing is significantly reduced due to minimized KV cache (only 1/6 of Hy4’s size).
- Native Multimodal Support: Unified handling of text, images, audio, and video—no need for additional models for vision or audio understanding tasks.
- Superior Daily Task Performance: On internal WorkIndex benchmark, overall effectiveness slightly surpasses Hy4 preview, with user satisfaction approaching that of Claude Haiku 5.5.
- High Output Conciseness: Reinforcement learning reduces end-to-end output length by 50%, resulting in more concise responses with clearer decision-making rationale.
Project Repository
- Official Website: https://hunyuan.tencent.com/research/hy4-flash
Competitive Comparison with Similar Models
| Comparison Dimension |
Hy4 flash |
DeepSeek-V4.1-Flash |
| Model Architecture |
224B total / 12B activated MoE with 64B Engram |
MoE architecture, exact parameters undisclosed |
| Context Length |
1M tokens |
1M tokens |
| Multimodal Support |
Native support for text, image, audio, video |
Text-focused; weak multimodal capabilities |
| Decoding Speed |
Over 150 tps |
Approx. 100 tps |
| API Pricing |
Input: 0.5 RMB, Output: 2 RMB/million tokens (50% discount until end of November) |
Peak-hour pricing ~8x higher than Hy4 flash |
| Output Conciseness |
50% reduction in output length after RL optimization (29% shorter than DeepSeek in WeChat Pay scenarios) |
Relatively verbose outputs |
| Typical Task Cost |
Single-task cost on WorkIndex ≈ 1/8 of competitor |
Baseline level |
| Ecosystem Integration |
Tencent ecosystem (Yuanbao, WorkBuddy, QQ, ima, etc.) + TokenHub, soon to join OpenRouter |
DeepSeek official platform and third-party APIs |
Application Scenarios
- Office Document & PPT Processing: Summarization, editing, table repair, and automated generation/modification of PPTs—ideal for frequent workplace needs.
- Deep Search & Information Aggregation: Low-cost execution of multi-round retrieval, cross-validation, and structured report generation with sources—suited for research and competitive analysis tasks.
- Enterprise AI Assistant: Autonomous planning of tool calls within platforms like Enterprise WeChat and ima to perform information retrieval, document analysis, and knowledge base Q&A.
- Computer Use Automation: Executes step-by-step GUI operations such as app downloads/installations, system setup/diagnostics, and payment client flows—already applied in Marvis and WeChat Pay scenarios.
Source: AI Tools Hub