Hunyuan Hy3 API

Hunyuan Hy3 is a general-purpose large language model officially released and open-sourced by Tencent's Hunyuan team on July 6, 2026. It uses a fast-slow thinking fusion MoE architecture (total parameters 295B, activated parameters 21B), supports 256K context length, and strengthens Coding, long-form writing, complex reasoning, and Agent capabilities. In Tencent's internal 270-expert blind test, it scored 2.67/4, outperforming GLM 5.1's 2.51/4, especially in front-end design and data-processing productivity tasks.

Model Positioning

Hunyuan Hy3 is positioned as a high-value, practical general-purpose LLM that matches the performance of much larger flagship models with a smaller parameter footprint (21B activated). It is the production version following Hunyuan's infrastructure rebuild; compared to Hy3 Preview (released April 23, 2026), it leaps forward again in reasoning, instruction following, code generation, and agentic capabilities — Agent task success rate rose from 72% to 90% and hallucination rate dropped by more than half. It suits Agent automation, Coding, long-form writing, complex reasoning, and cost-sensitive high-frequency workloads.

Core Capabilities

Fast-slow thinking fusion

Enable thinking mode via thinking: {"type": "enabled"} and control reasoning depth with reasoning_effort (low/medium/high) — save latency and cost on simple tasks, get deep reasoning on complex ones.

256K long context

256K-token context window (max input 192K / max output 128K), suitable for long-document analysis, knowledge-base QA, and repository-level code understanding.

Strong Agent capability

Agent task success rate improved to 90% (Hy3 Preview: 72%), hallucination rate dropped by more than half — suitable for autonomous multi-step workflows: planning, research, spreadsheet work, report writing.

Coding, long-form, and reasoning enhancement

Post-trained on real business scenarios; excels in code generation, long-form writing, and complex logical reasoning, supporting cross-file refactoring, long-document analysis, and complex multi-turn Agent interaction.

Cached-read discount

Cached reads billed at $0.05/million tokens (one quarter of the standard $0.20/million input price), significantly reducing costs for repeated-context workloads.

OpenAI-compatible interface

Compatible with OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages API protocols; callable with any compatible SDK (Python, Node.js, etc.).

Pricing

Total cost = input tokens × $0.20/million + output tokens × $0.80/million (cached reads at $0.05/million).

Note: Tencent's official TokenHub pricing is ¥1/million input, ¥4/million output, ¥0.25/million cache hit (CNY); the USD prices above are iCreat platform prices.

Use Cases

• Agent automation: multi-step autonomous workflows — planning, research, spreadsheets, reports — 90% task success rate
• Coding: code generation, debugging, refactoring, and cross-file understanding in common languages
• Long-form and knowledge-base writing: 256K context supports long-document drafting based on personal knowledge bases and reference materials
• Complex reasoning: data extraction, invoice processing, structured reports, and other logic-intensive tasks
• Cost-sensitive high-frequency calls: cached reads at only $0.05/million tokens — ideal for repeated-context workloads (RAG, multi-turn conversations with shared system prompts)

Model Comparison

Hunyuan Hy3 vs. Hy3 Preview

Hunyuan Hy3 vs. Hunyuan T1 and Hunyuan TurboS

Note: Hunyuan T1 and TurboS were taken offline on June 22, 2026; Hy3 is Hunyuan's current flagship model.

Why Choose Hunyuan Hy3?

• High value: 21B activated parameters match much larger flagships; input at only $0.20/million tokens
• 90% Agent success rate: reliable multi-step autonomous workflows; hallucination rate cut by more than half
• Fast-slow thinking fusion: one model covers simple chat and deep reasoning — no model switching needed
• 256K long context: supports repository-level code understanding and long-document analysis
• Apache 2.0 open source: self-deployable, commercially unrestricted, already adapted to multiple domestic chips
• Cached reads at $0.05/million: dramatically lowers costs for repeated-context workloads

Specifications

Architecture

Hunyuan Hy3 uses a Mixture-of-Experts (MoE) architecture with 295B total parameters and 21B activated per token, supporting 256K context length. The model fuses fast and slow thinking: fast responses by default, with deep reasoning enabled via the thinking parameter. Post-training optimizes data quality and diversity based on real business scenarios, leaping forward again in reasoning, instruction following, code generation, and agentic capabilities over Hy3 Preview.

Tencent Hunyuan launched an infrastructure rebuild in late January 2026, released Hy3 Preview on April 23 (the first model after the rebuild), and officially released Hy3 on July 6. The model is open-sourced under Apache 2.0 and has been adapted to hardware platforms including DeepVerse (Skywalker), Hygon DCU, and Intel Xeon 6 and Gaudi 2. An extreme quantization version was released on July 14, further lowering deployment barriers.

Notes

• Text-only model: does not support image, audio, or video input
• reasoning_effort only accepts fixed values (low/medium/high); unsupported values may be ignored
• Cached-read discount applies only when the input prefix hits the model cache
• The 256K context window is a shared request space; reserve room for system instructions, conversation state, and the final answer
• When thinking mode is enabled, responses include reasoning token consumption; for latency- and cost-sensitive high-throughput scenarios, use reasoning_effort: "low"
• On non-Coding / non-reasoning creative tasks, performance may trail specialized flagship models

FAQ

What is the difference between Hunyuan Hy3 and Hy3 Preview?

Hy3 is the production version with stronger reasoning, Coding, and Agent capabilities (90% vs. 72% task success rate) and a halved-plus hallucination rate. Both share the same architecture (MoE 295B/21B, 256K context); Hy3 is positioned as a production-grade high-value model, while Hy3 Preview is an early public version.

How do I enable thinking mode?

Set "thinking": {"type": "enabled"} in the request body, optionally adjusting reasoning_effort (low/medium/high). Lower or disable thinking on simple tasks to save latency and cost; raise the tier for complex tasks.

What is Hy3's context window?

256K tokens — max input 192K, max output 128K.

How is cached-read billing calculated?

Cached input is billed at $0.05/million tokens instead of the standard $0.20/million input price. Ideal for RAG and multi-turn conversations with shared system prompts.

Can I call it directly with the OpenAI SDK?

Yes. Set base_url to https://api.icreat.ai/llm/openai/v1 and api_key to your iCreat API Key — no extra adaptation needed. temperature, max_tokens, top_p, and other standard parameters are all available.

Can I self-host?

Yes. The model weights are open-sourced under Apache 2.0, already adapted to multiple domestic chips (DeepVerse, Hygon DCU) and Intel hardware platforms, and deployable via vLLM and similar inference frameworks, commercially unrestricted. An extreme quantization version was released on July 14, further lowering deployment barriers.