Hunyuan 3
#42 in Multimodal Modelstencent · v3 · seit 2026-04-23 · 2× · zuletzt 30. Juni 2026
10
Momentum
Hunyuan 3 (also known as "Hy3 Preview") is Tencent's currently most capable open-weight language model, released on April 23, 2026. It is based on a Mixture-of-Experts architecture with 295 billion total parameters (21 billion active per forward pass) and an additional 3.8 billion MTP layer for speculative decoding. The model is the first trained on Tencent's rebuilt pretraining and reinforcement learning infrastructure, focusing on complex reasoning, coding, agent tasks, and instruction following. At time of release, it is a text-only model with no documented native image or video input capability.
Momentum trend
04.04.03.07.
Features
| Context Window (Tokens) | 256,000 tokens (256K); some sources also cite 262,144 tokens (262K) |
| Multimodal Inputs | Text only (language/code); no confirmed native image or video inputs for Hy3 Preview – the 'Preview' label indicates multimodal variants are still pending |
| On-Device vs. Cloud | Cloud (Tencent Cloud TokenHub, OpenRouter); open-weight model also available for self-hosted deployment on 2x H100 80GB or 3x A100 80GB |
| Price per Unit | From CNY 1.2 / million input tokens and CNY 4.0 / million output tokens via Tencent Cloud TokenHub; internationally approx. $0.07 / million input tokens and $0.26 / million output tokens via OpenRouter |
| Vision-Language Benchmark Score | No confirmed vision-language benchmark score found – Hy3 Preview is a text-only model at release with no documented VLM benchmarks |
