Hy3
Completed Hy3: same 295B/21B MoE, now Apache-2.0, API input at 1 yuan per million tokens; scored 2.67 vs GLM-5.1's 2.51 in a 270-expert blind test.
Post-training fed by feedback from over 50 Tencent products; Tencent reports the hallucination rate fell from 12.5% to 5.4% on internal evals. Built into Yuanbao, WorkBuddy and others.
- Date
- Monday, 6 July 2026
- Lab
- Tencent
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-Bench Pro | 57.9 vs Claude Opus 4.8 69.2 | company |
| Blind expert eval (270 experts, scale of 4) | 2.67 vs GLM-5.1 2.51 Tencent-run | company |
| Hallucination rate (internal) | 12.5% to 5.4% | company |
| API price (input / output) | 1 yuan / 4 yuan per M tokens | company |
Output price and benchmark breakdown come from BigGo's write-up; Caixin confirms the 1 yuan input price. Benchmarks are company-run.
Sources
- www.tencent.com/en-us/articles/2202386.html
- huggingface.co/tencent/Hy3
- www.caixinglobal.com/2026-07-06/tencent-launches-upgraded-hunyuan-3-ai-model-with-free-age
- finance.biggo.com/news/eac480c0-1840-4271-858a-eb43389b8811
This record was checked against its sources on 6 October 2026. How we check