Hy3 – Tencent Hunyuan's Open-Source Mixture of Experts Model

1. What is Hy3

Hy3 is a 295B-parameter Mixture of Experts (MoE) model open-sourced by the Tencent Hunyuan team. It demonstrates significant improvements in agent capabilities, reasoning, and long-context tasks, with multiple benchmarks rivaling flagship models that are 2–5 times larger in parameter count. Through enhanced post-training and fine-grained data cleaning, the model's hallucination rate has been reduced to 5.4%, common-sense error rate to 12.7%, and multi-turn question error rate to 7.9%. Hy3 supports diverse productivity scenarios such as code development, office automation, and front-end design, and offers a highly cost-effective and reliable choice with extremely low API costs (input: 1 RMB per million tokens).

Image source: Official article
Image source: official article

Technical Positioning and Domain: Hy3 belongs to the domain of large-scale Mixture of Experts (MoE) language models, focusing on agent tasks, code development, office automation, and other productivity scenarios. Its unique positioning lies in achieving performance that surpasses larger dense models or larger MoE models with only 21B activated parameters, while maintaining low inference costs, filling the market gap for "high efficiency, low price" large models.

Research Background: Developed by the Tencent Hunyuan team, building on experience from the previous preview version, Hy3 systematically improves the quality and diversity of post-training data and expands the scale of reinforcement learning computation, significantly enhancing model capabilities. The team has accumulated deep technical expertise in data cleaning and software-hardware co-optimization, making Hy3 an important milestone in the Tencent Hunyuan series.

Core Value: Hy3 addresses key pain points in real-world applications of large models, such as high costs, high hallucination rates, and unstable tool invocation. Released under the open-source Apache 2.0 license, Hy3 matches or even surpasses flagship models 2–5 times its parameter size on benchmarks like SWE-bench Verified, Terminal Bench, and BrowseComp, providing enterprises and developers with a viable solution of "leapfrog performance at extremely low prices."

Technical Features: Hy3 adopts an MoE architecture with 295B total parameters but only 21B activated, achieving extremely high inference efficiency. Joint training optimizes coreference resolution and multi-turn constraint inheritance, significantly improving long-range interaction consistency. Software-hardware co-optimization reduces inference costs, with office task token consumption approximately 47–49% lower than competitors. Fine-grained data cleaning follows the principle of "only answer when there is evidence," effectively suppressing hallucinations.

2. Key Features

• Agent Tasks: Significant improvements in reasoning, agent, and long-context tasks, with multiple benchmarks matching flagship models 2-5 times its parameter size. For example, it achieves 84.2 on BrowseComp, 79.1 on MCP Atlas (public), 68.5 on ClawEval (pass**3), 55.3 on SkillsBench (text-only), 53.2 on HLE (with tools, text-only), and 73.4 on AA-LCR, all outperforming or matching larger models like DeepSeek-V4 pro.
• Code Development: Supports the full workflow of software development, frontend design, and CI/CD. It scores 78.0 on SWE-bench Verified and 71.7 on Terminal Bench 2.1, and can generate directly runnable HTML web pages and browser plugins, significantly boosting development efficiency.
• Office Productivity: Supports Excel modeling and analysis, PPT creation, complex data aggregation, and linked formula calculations. Token consumption for office tasks is reduced by approximately 47-49% compared to competitors. Combined with extremely low API pricing (output: 4 yuan per million tokens), it drastically lowers the cost of enterprise office automation.
• Tool Calling: Enhanced tool calling stability and cross-framework generalization, supporting mainstream frameworks like Codebuddy, Cline, and KiloCode. The standard deviation of SWE Bench Verified scores across different frameworks is controlled within 4 percentage points. Tool call error recovery capabilities are significantly improved, ensuring reliability in production environments.
• Anti-Hallucination: Through fine-grained data cleaning and training constraints, it adheres to the principle of "answer only with evidence, explicitly state missing information." The model hallucination rate has dropped from 12.5% to 5.4%, the common sense error rate from 25.4% to 12.7%, and the multi-turn question rate from 17.4% to 7.9%, making it reliable in scenarios requiring high information accuracy.
• Multi-Turn Dialogue: Optimizes anaphora resolution, ellipsis recovery, and multi-turn constraint inheritance to maintain intent consistency in complex, long-range interactions. In scenarios requiring sustained context, such as customer service and AI avatars, it can handle incomplete expressions and make reasonable judgments based on context, avoiding irrelevant responses.

3. How to Use

1. Environment Requirements: It is recommended to use an NVIDIA A100 (80GB) or higher GPU, with Linux (Ubuntu 20.0+) as the operating system, and install Python 3.8+, PyTorch 2.0+, and the transformers library. If using only the API, no local GPU is needed; only a Tencent Cloud account is required.
2. Obtain Model Weights: Visit the GitHub repository (link to be updated after official release) to clone or directly download the model weights and inference code. The model has approximately 295B parameters, so ensure sufficient disk space (at least 600GB).
3. Activate API Service: In the Tencent Cloud console, search for "Hunyuan Large Model" or "Hy3" to activate the API service and obtain access keys (SecretId and SecretKey). The API input price is 1 yuan per million tokens, output is 4 yuan per million tokens, and cache-hit input is only 0.25 yuan, offering excellent cost-effectiveness.
4. Configure Runtime Environment: Set up the runtime environment locally or in the cloud, load the model, and integrate it with scaffolding tools such as Codebuddy, Cline, or KiloCode. Use the official Python script for model loading and inference, or call the API via an HTTP interface.
5. Call to Complete Tasks: Invoke the model via API or local deployment, and write prompts based on the task type (code development, office automation, data analysis, etc.). It is recommended to use the official system prompt template and enable tool-calling mode for optimal performance.
6. Notes: When deploying locally, pay attention to GPU memory usage; consider using quantization or inference frameworks like vLLM for optimization. When using the API, be mindful of rate limits and batch process requests to reduce costs.

4. Pros and Cons Analysis

5. Comparison with Similar Tools

Selection Recommendations: For teams pursuing top-tier performance with ample budgets, DeepSeek-V4 pro slightly edges ahead on benchmarks like SWE-bench, but comes with higher API costs and larger activated parameters. For enterprises prioritizing cost-effectiveness and low hallucination requirements, Hy3 delivers near or even superior intelligence at an extremely low price, making it especially suitable for high-frequency use cases such as office automation and code development. For users needing a closed-source ecosystem and the strongest multimodal capabilities, GPT-4o remains the benchmark, though it is costly and cannot be deployed locally. Overall, Hy3 offers the best cost-performance ratio among open-source MoE models, making it ideal for small to medium-sized teams and cost-sensitive projects.

6. Editor's Take

Hy3's technological innovation is reflected in three aspects: first, its MoE architecture with 295B total parameters and 21B activated parameters achieves performance surpassing that of larger models, validating the feasibility of the "small activation, large capacity" approach; second, through joint training and fine-grained data cleaning, it reduces the hallucination rate to 5.4% and the common sense error rate to 12.7%, reaching industry-leading reliability; third, software-hardware co-optimization reduces token consumption for office tasks by nearly 50%, offering significant advantages in cost control.

In terms of practical value, Hy3's extremely low API price (1 RMB per million tokens for input) and high reliability make it an ideal choice for enterprise-level AI deployment, especially suitable for high-frequency, high-accuracy production scenarios such as code development, office automation, and intelligent customer service. The open-source Apache 2.0 license further lowers the barrier to entry, allowing commercial customization and secondary development.

Target users include: independent developers in need of a low-cost, high-performance AI assistant; enterprise IT teams seeking cost reduction and efficiency improvement; and academic institutions researching MoE architectures. With significant future development potential, leveraging the Tencent Hunyuan ecosystem, optimized versions for more vertical domains are expected, along with community-driven improvements to the toolchain.

. The 0.5-star deduction is due to unverified performance in Chinese scenarios and incomplete documentation and community support, but overall performance and cost have exceeded expectations.

7. Application Scenarios

• Software Development: Supports code generation, debugging, and CI/CD workflow automation, scoring 78.0 on SWE-bench Verified. It can be integrated with scaffolding tools like Codebuddy and Cline, helping developers quickly complete feature implementation, bug fixes, and code reviews, significantly improving development efficiency.
• Office Automation: Supports Excel modeling and analysis, PPT creation, complex data aggregation, and linked formula calculations. Token consumption is 47-49% lower than competitors. Combined with extremely low API pricing, enterprises can embed Hy3 into internal OA systems to automate repetitive tasks such as report generation, data cleaning, and meeting minutes.
• Frontend Design: Can generate directly runnable HTML pages and browser plugins, supporting UI generation directly from natural language descriptions. Designers and frontend developers can use Hy3 for rapid prototyping, reducing manual coding time.
• Intelligent Customer Service: Supports AI avatars and customer service scenarios, capable of handling incomplete expressions and making reasonable judgments based on context. The multi-turn question rate is only 7.9%, with a hallucination rate of 5.4%. It can replace human agents for common inquiries, reducing enterprise labor costs.
• Game Development: Supports game logic writing and integration with game assistants, such as the AI assistant for Path of Exile: Descent. Game developers can use Hy3 to generate NPC dialogue, quest scripts, and balance adjustment suggestions, accelerating the game development cycle.

8. FAQ

Q: Is Hy3 fully open-source? Can it be used commercially?
A: Yes, Hy3 is licensed under Apache 2.0. The code and model weights are open-sourced on GitHub and HuggingFace, allowing commercial use, modification, and redistribution without additional authorization.

Q: What hardware is required for local deployment?
A: It is recommended to use NVIDIA A100 (80GB) or higher-end GPUs, with a minimum VRAM requirement of 80GB. With optimizations like quantization or vLLM, it can run on 48GB VRAM (e.g., A100 40GB), but performance may be affected. For individual users, it is recommended to use the Tencent Cloud API directly.

Q: What is the API pricing, and how do I activate it?
A: API input costs 1 RMB per million tokens, output costs 4 RMB per million tokens, and cache-hit input costs only 0.25 RMB. Search for "Hunyuan Large Model" or "Hy3" in the Tencent Cloud console to activate it, then obtain the key and call it via the HTTP interface.

Q: How is the model's hallucination rate controlled? Does it support Chinese?
A: Through fine-grained data cleaning and training constraints, following the principle of "only answer with evidence," the hallucination rate is reduced to 5.4%, and the common error rate is reduced to 12.7%. The official documentation does not specify Chinese performance, but based on general English training, Chinese tasks may not perform as well as models optimized specifically for Chinese. It is recommended to test before use.

Q: How can Hy3 be integrated into existing tools?
A: It supports mainstream scaffolding tools like Codebuddy, Cline, and KiloCode, and can be accessed via API or local deployment. The official Python SDK and sample code are provided, and developers can refer to the documentation in the GitHub repository for quick integration.

Q: What is the model's context length? Does it support long texts?
A: The official documentation does not specify the exact context length, but based on performance in long-context task benchmarks (e.g., AA-LCR 73.4), it supports processing relatively long contexts. For practical use, it is recommended to keep inputs within 8K-16K tokens for optimal results.

9. Project Address

• Project Website: https://hy.tencent.com/research/hy3
• GitHub Repository: https://github.com/Tencent-Hunyuan/Hy3
• HuggingFace Model Library: https://huggingface.co/tencent/Hy3
• Tencent Hunyuan Official Organization: https://github.com/Tencent-Hunyuan