Overview
- Tool name: Hunyuan Hy3 (Hunyuan 3.0)
- Developer: Tencent (Hunyuan team, led by Chief AI Scientist Shunyu Yao)
- Official website: https://cloud.tencent.com/product/hunyuan
- API: Tencent Cloud's TokenHub model platform
- Category: AI Models
- License: Apache 2.0 (open weights)
Hunyuan Hy3 is Tencent's next-generation model, officially released and open-sourced on July 6, 2026 - the second version after the team rebuilt its foundation infrastructure from scratch in January 2026 (the first being April's Hy3 preview). Key traits:
- A fast/slow-thinking fused MoE: 295B total parameters with only 21B active, 256K context
- Small frames, flagship scores: officially "matching flagships 2-5x its size", beating GLM 5.1 in a 270-expert blind test (2.67 vs 2.51 out of 4)
- Battle-tested: task success on WorkBuddy jumped from 72% to 90% with average time down 34%; daily token consumption grew 20x since the preview
Hy3 powers Tencent's product line (WorkBuddy, CodeBuddy, Yuanbao, ima, Marvis and more), is listed on OpenRouter and other overseas platforms, and ships an aggressively quantized build plus adaptations for Intel Xeon/Gaudi and domestic accelerators.
API pricing (official, per million tokens)
- Input CNY 1 / output CNY 4, cache-hit input CNY 0.25
- Agent-oriented Token Plans start at CNY 28/month for individuals
Pricing verified against Tencent's official launch announcements and Tencent Cloud TokenHub (October 2026).
Who it is for
- Productivity tasks: software development, office work (Tencent Docs AI PPT success +20%), financial modeling, front-end design, game production
- Agent development: optimized for multi-turn constraint inheritance and coreference resolution, stable long-horizon tool calls
- Cost-critical production: at CNY 1/4, among the cheapest in its open-flagship class
- Sovereign deployment: Apache 2.0 with no regional restrictions plus Ascend/Hygon/Intel adaptations
Limitations
- 256K context is smaller than the 1M windows of DeepSeek V4 / Kimi K3
- A pure language model; multimodal lives on separate Hunyuan product lines (image/video)
- Less general-benchmark visibility than Qwen/DeepSeek and a thinner English-community knowledge base
Frequently asked questions
What is the latest Hunyuan model?
Hy3 (GA July 6, 2026): 295B total / 21B active, 256K context, Apache 2.0 open weights.
What does the Hy3 API cost?
Official pricing is CNY 1 input / CNY 4 output per million tokens, cache-hit input CNY 0.25 (Tencent Cloud TokenHub).
Is it open source?
Yes - Apache 2.0, free for commercial use worldwide, downloadable from HuggingFace and ModelScope.
Hy3 vs Qwen3.5 vs DeepSeek V4?
All in China's open-source first tier: Hy3 wins on agent stability and the rock-bottom CNY 1/4 price; Qwen3.5 on native multimodality and its full size family; DeepSeek V4 on 1M context and international community ecosystem.
*This page is not affiliated with Tencent. Benchmarks are vendor-reported; pricing per Tencent Cloud TokenHub. Verified as of October 2026.*