Tencent Hy3

What it is

Tencent Hy3, also called Hunyuan Hy3, is an open-weight model that combines fast and slow thinking. Tencent released it on 6 July 2026.^[3] It is a mixture-of-experts (MoE) model with 295 billion total parameters, of which 21 billion are activated per token.^[1] Its context length is up to 256K tokens.^[1] Hy3 is the successor to HY 2.0, which had 406 billion total parameters.^[1][3]

Origins

Tencent first open-sourced a preview of the model on 23 April 2026. That preview was released under the Tencent Hy Community License, and the final Hy3 was released under the Apache License 2.0, with an FP8 version also available.^[1][2] Apache 2.0 is a permissive open-source license that allows use, modification and redistribution.

The preview's base model, before instruction tuning, scored 87.42 on MMLU, 65.76 on MMLU-Pro, 95.37 on GSM8K and 89.80 on C-Eval.^[2] These are the base-model figures Tencent published for the preview.

Main features

The model uses 192 experts with top-8 routing, meaning that for each token the router selects eight of the experts, and it has 80 layers.^[1] It also includes a multi-token-prediction layer of 3.8 billion parameters.^[1] Multi-token prediction trains the model to predict several upcoming tokens at once, which can be used to speed up generation. Attention uses 64 heads with grouped-query attention, using 8 key-value heads, and the vocabulary has 120,832 tokens.^[1]

Reasoning effort can be set to no_think, which is the default, low or high.^[1] This lets a user choose between a direct answer and longer reasoning on harder problems.

Tencent reported the following scores: 90.4 on GPQA Diamond, 78 on SWE-bench Verified, 57.9 on SWE-bench Pro, 71.7 on Terminal-bench 2.1 and 75.8 on SWE-bench Multilingual.^[1] In a blind evaluation by 270 experts, Hy3 averaged 2.67 out of 4, against 2.51 for GLM-5.1.^[1] Tencent's internal evaluation reported that the hallucination rate fell from 12.5% to 5.4%, and the multi-turn issue rate fell from 17.4% to 7.9%.^[1]

Pricing on Tencent Cloud TokenHub was RMB 1 per million input tokens, RMB 4 per million output tokens and RMB 0.25 per million cached input tokens.^[3] Hy3 has been integrated into WorkBuddy and CodeBuddy, Yuanbao, Marvis and ima.^[3]

Place in the field

Hy3 was the first major release under Yao Shunyu, who leads Tencent's LLM department.^[4] Tencent's own benchmarks placed the model near midsize models such as GLM-5.2 and DeepSeek V4 Pro.^[4] Its release was followed less than two months later by the larger Hy4 preview.