Hy3 is a large language model that combines reasoning and agent capabilities with cost efficiency through a Mixture-of-Experts architecture.
Hy3 addresses the need for capable reasoning and agentic models that remain cost-effective to deploy. The model uses a 295-billion-parameter MoE architecture with only 21 billion active parameters, reducing computational overhead while maintaining performance. It incorporates a specialized MTP layer and supports a 256K context length, enabling it to handle long-form reasoning and complex multi-step tasks. The model underwent extensive post-training with higher-quality data and reinforcement learning to improve performance across reasoning, agentic, and long-context scenarios.
Developers should consider Hy3 for productivity-focused applications including coding, office work, financial modeling, frontend design, and game development where cost efficiency matters alongside capability. The model performs competitively with much larger flagship models while requiring significantly fewer parameters. A key distinguishing factor is its production-grade reliability for tool calling and output format constraints, addressing real-world deployment concerns. The model generalizes well across different agent scaffolding frameworks, with accuracy variance remaining within four percentage points across different implementations. Blind evaluation with domain experts showed the model outperforming comparable alternatives in frontend development, data and storage tasks, and CI/CD workflows.
The project provides comprehensive deployment documentation covering vLLM and SGLang integration, finetuning capabilities, reinforcement learning post-training procedures, and quantization options. Development activity shows sustained focus on production reliability, with specific attention to fixing tool-call stability issues and output format consistency based on feedback from deployed products. The team maintains active engagement with real-world use cases, incorporating learnings from over fifty products into model improvements.
