Seed Thinking v1.5
FreeAdvancing Superb Reasoning Models with Reinforcement Learning
FreeFree tier
Inputs: textOutputs: text
About Seed Thinking v1.5
Seed-Thinking-v1.5 is a Mixture-of-Experts (MoE) reasoning model developed by ByteDance, designed to think before responding, leading to improved performance across a variety of benchmarks. It achieves 86.7 on AIME 2024, 55.0 on Codeforces, and 77.3 on GPQA, demonstrating strong reasoning capabilities in STEM and coding. With 20 billion activated parameters (200 billion total), it surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, showing broad generalization. The model is open-source and accompanied by a technical report and internal benchmarks (BeyondAIME and Codeforces) to be released.
Key Features
Reasoning through thinking before responding improves benchmark performance
Mixture-of-Experts (MoE) architecture with 20B activated and 200B total parameters
Achieves 86.7 on AIME 2024, 55.0 on Codeforces, and 77.3 on GPQA
Surpasses DeepSeek R1 by 8% win rate on non-reasoning tasks
Open-source release with technical report and internal benchmarks (BeyondAIME, Codeforces)
Pros & Cons
Pros
- State-of-the-art reasoning performance on STEM benchmarks, especially mathematics and coding
- Efficient MoE design with only 20B activated parameters
- Generalization to non-reasoning domains, outperforming larger models like DeepSeek R1
- Open-source and freely available for research and development
Cons
- Lower factuality scores (SimpleQA 12.9%) compared to some models like DeepSeek R1 (30.1%)
- Not top-performing in all categories (e.g., SWE-bench, Aider Polyglot) against larger models
- Relatively new and less mature ecosystem compared to established reasoning models
Best For
Mathematical problem-solving and competition-level reasoningCoding challenges and competitive programmingScientific question answering and research assistanceGeneral knowledge tasks and instruction followingAgentic coding and software engineering tasks
FAQ
What is Seed-Thinking-v1.5?
Seed-Thinking-v1.5 is an open-source Mixture-of-Experts reasoning model developed by ByteDance that uses reinforcement learning to think before responding, improving performance on a wide range of tasks.
What benchmarks does Seed-Thinking-v1.5 excel at?
It achieves high scores on AIME 2024 (86.7), Codeforces pass@8 (55.0), and GPQA diamond (77.3), among others.
How does Seed-Thinking-v1.5 compare to DeepSeek R1?
Seed-Thinking-v1.5 achieves an 8% higher win rate on non-reasoning tasks while being a smaller MoE model (20B activated vs. DeepSeek R1's full parameters).
Is Seed-Thinking-v1.5 open source?
Yes, the model and a technical report are publicly available on the ByteDance-Seed GitHub repository.