Seed1.5-Thinking
200B-total, 20B-active MoE reasoning model trained with RL, reported 86.7 on AIME 2024 and 8% win-rate edge over DeepSeek-R1 on non-reasoning tasks.
Report details the RL recipe and releases two internal benchmarks, BeyondAIME and a Codeforces set. Served on Volcano Engine as a Doubao thinking model; weights not released.
- Date
- Thursday, 10 April 2025
- Lab
- ByteDance Seed
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| AIME 2024 | 86.7 | company |
| Codeforces (internal metric) | 55.0 | company |
| GPQA | 77.3 | company |
| Total / active parameters | 200B / 20B | company |
Date is arXiv v1; the commercial Doubao-1.5-thinking-pro launch was days later (Volcano Engine model ID 250415 suggests mid-April). Results are ByteDance-run.
Sources
This record was checked against its sources on 6 October 2026. How we check