Hunyuan-T1 (on TurboS)
Reasoning model built on TurboS, the first ultra-large hybrid Transformer-Mamba MoE; 96.7% of post-training compute went to RL and MMLU-Pro reached 87.2.
TurboS (560B total, 56B active, 16T tokens, 256K context, adaptive long-short CoT) was disclosed in the May 2025 paper and ranked top 7 on LMSYS Arena with 1356. T1 claims parity with DeepSeek-R1 and decoding speed 2x faster under the same deployment.
- Date
- Friday, 21 March 2025
- Lab
- Tencent
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| MMLU-Pro | 87.2 second to o1 per Tencent | company |
| GPQA-Diamond | 69.3 | company |
| MATH-500 | 96.2 | company |
| Share of post-training compute spent on RL | 96.7% | company |
Official-version date 2025-03-21 is from memory of the launch and not stated in the README I read (it says TurboS came 'at the beginning of March' and a T1-Preview in mid-February). Treat the day as unconfirmed.
Sources
This record was partly confirmed: some claims could not be checked on 6 October 2026. How we check