Hy3

Overview

Hy3 is a text foundation model from Tencent Hunyuan, released on July 6, 2026. It is available through the API with the model ID hy3. The model uses a Mixture-of-Experts (MoE) architecture with 295 billion total parameters and 21 billion activated parameters, and is designed for coding, long-context knowledge work, reasoning, and tool-driven agent workflows.

Key Features

• Hybrid Reasoning: Supports direct responses and deeper reasoning. The documented modes are no_think, think_low, and think_high, allowing developers to balance response speed with reasoning depth.
• Sparse MoE Architecture: Uses 295 billion total parameters and 21 billion activated parameters. Its architecture includes 192 experts, with eight activated per token, plus 3.8 billion MTP-layer parameters.
• Long-Context Generation: Supports a 256K-token context window, a 192K-token maximum input, and a 128K-token maximum output.
• Agent-Oriented API Support: Supports structured output, function calling, context caching, and preserved reasoning for tool-driven workflows.

Best Use Cases

• Software Engineering Agents: Repository-scale implementation, debugging, frontend development, CI/CD work, and multi-step tool use.
• Office and Knowledge Work: Long-document analysis, report and presentation preparation, knowledge-base question answering, and structured file generation.
• Financial and Data Workflows: Financial modeling, data analysis, and other workflows that combine reasoning with external tools and iterative verification.
• Long-Horizon Agent Tasks: Workflows that require planning, repeated tool calls, multi-turn constraint retention, and recovery from tool-call errors.
• Self-Hosted Model Customization: Private deployment, fine-tuning, quantization, or reinforcement-learning post-training with the available open weights.

Capabilities and Limitations

Capability	Description
Reasoning	Supports direct and deeper reasoning modes: no_think, think_low, and think_high.
Creative Writing	Supports general text generation and long-form content workflows.
Coding	Designed for coding agents and software-development workflows. Tencent reports particularly strong results in frontend development, data and storage, and CI/CD tasks.
Multimodal	Text input and text output only.
Context Window	256K tokens.
Maximum Input	192K tokens.
Max Output	128K tokens.
Tool Use	Supports function calling, structured output, automatic tool selection, and preserved reasoning across tool calls.

Known Limitations

• Hy3 is a text-only model; applications that require native image, audio, or video understanding need a separate multimodal model.
• When preserved reasoning is enabled for multi-turn tool use, clients must return the assistant message's reasoning_content together with its tool calls and tool results. Dropping that state interrupts the documented continuation workflow.

Pricing

Model	Input (Credits/Token)	Cache Write (Credits/Token)	Cache Read (Credits/Token)	Output (Credits/Token)	Web Search (Credits/Use)
Hy3	0.132	0.132	0.033	0.528	-

Pricing note

Prices shown in the documentation are B.AI standard reference prices for base billing purposes. B.AI may provide lower actual usage costs through top-up bonuses and account benefits. Specific prices, bonus Credits, and account benefits are subject to the platform display and final billing records.