tencent/Hy3
Hy3 is a 295 billion parameter Mixture-of-Experts (MoE) model developed by the Tencent Hy Team, featuring 21 billion active parameters and a 256K context length. This model is designed for robust agentic capabilities and reliable product experiences, demonstrating significant improvements in tool call stability, hallucination reduction, and complex context retention. It is optimized for productivity scenarios such as coding, office work, and financial modeling, offering a cost-effective option competitive with larger flagship models.
Loading preview...
Overview
Hy3 is a 295 billion parameter Mixture-of-Experts (MoE) model from the Tencent Hy Team, featuring 21 billion active parameters and a 256K context length. It utilizes 192 experts with 8 activated, and supports BF16 precision. The model has been scaled up with higher quality data and advanced RL training, aiming to outperform similar-sized models and rival larger flagship open-source models in performance and utility.
Key Capabilities
- Stronger Agent Capabilities: Demonstrates solid gains across reasoning, agentic, and long-context tasks, competitive with much larger models. It shows significant progress in productivity scenarios like coding, office work, financial modeling, and game development.
- Enhanced Reliability: Addresses common issues identified through product feedback, including improved stability of tool calls and output formats, with tool-call error recovery and efficiency. It generalizes across different agent scaffoldings, maintaining accuracy variance within 4% on SWE-Bench Verified.
- Reduced Hallucination: Implemented fine-grained data cleaning and training constraints, reducing the hallucination rate from 12.5% to 5.4% and commonsense error rates from 25.4% to 12.7% in internal evaluations.
- Complex Context Retention: Improved coreference resolution, ellipsis recovery, and multi-turn constraint inheritance, reducing the issue rate in multi-turn tests from 17.4% to 7.9% and enhancing long-dialogue evaluations.
Deployment and Customization
- Supports deployment with vLLM and SGLang, with specific recipes available for MTP (Mixture-of-Thought) enabled serving.
- Provides a complete model finetuning pipeline and supports GRPO reinforcement learning training.
- Offers AngelSlim, a toolkit for large model compression, including quantization algorithms and speculative sampling.