Hy3

First flagship model built under Chief AI Scientist Yao Shunyu (former OpenAI researcher, core contributor to Operator and Deep Research). 295B total / 21B active MoE (192 experts, top-8, 80 layers, GQA with 64 query + 8 KV heads) with 1 MTP layer (3.8B params) for speculative decoding. 256K context. Developed in under three months following Tencent's March 2026 AI restructuring that merged AI Lab into the Hunyuan team.

SWE-bench Verified: 74.4, Terminal-Bench 2.0: 54.4, GPQA Diamond: 87.2. Base model: MMLU 87.42, GSM8K 95.37, SuperGPQA 51.60, LiveCodeBench-v6 34.86 (strongest in class on several). Deployed in Tencent's Yuanbao consumer app and CodeBuddy coding assistant. Designed to leverage WeChat's 1B+ users for contextual understanding. Preview weights shipped under the Tencent Hy Community License.

Official Hy3 launched July 6, 2026 (weights July 2), relicensed to Apache-2.0: same architecture, post-training scaled up with feedback from 50+ internal products. GA card reports SWE-bench Verified 78 (from 74.4), GPQA Diamond 90.4 (from 87.2), hallucination rate cut 12.5%→5.4%, and a 90% task-completion rate across Tencent's agent deployments. Integrated into WorkBuddy, CodeBuddy, Yuanbao, QQ Browser and ~50 more products; priced at ¥1/¥4 per M tokens on Tencent Cloud with a two-week free tier on OpenRouter. AA score (34) still reflects the April preview pending re-evaluation.