ByteDance has joined the rapidly evolving AI reasoning model competition, launching Seed-Thinking-v1.5, a new large language model focusing on Science, Technology, Engineering, and Mathematics (STEM) fields. This model, employing a Mixture-of-Experts (MoE) architecture, excels in various benchmark tests, even surpassing industry giants in certain metrics.
The Evolution of Reasoning AI
The reasoning AI race began in September 2024 with OpenAI's o1 model release, truly accelerating after DeepSeek R1's launch in January 2025. Major AI companies are now vying to develop models capable of "chain-of-thought" reasoning to provide more comprehensive and logical answers. Seed-Thinking-v1.5 utilizes the popular Mixture-of-Experts (MoE) architecture, similar to Meta's Llama4 and Mistral's Mixtral. This architecture allows the model to use only 20 billion parameters at a time from a vast 200-billion parameter pool, significantly improving efficiency.
Exceptional Performance
The model demonstrates impressive capabilities: achieving an 86.7% score on AIME2024, a 55.0% pass@8 score on Codeforces, and a 77.3% score on the GPQA science benchmark. Most notably, it surpasses Google's Gemini 2.5 Pro and OpenAI's o3-mini-high in the ARC-AGI benchmark. In non-reasoning tasks, Seed-Thinking-v1.5 boasts an 8.0% higher win rate than DeepSeek R1, indicating its performance advantage extends beyond logic or math-intensive tasks.
Technological Innovation and Breakthroughs
ByteDance employed several innovative techniques in developing Seed-Thinking-v1.5, including carefully curated training data, an advanced reinforcement learning framework, a dual-layer reward system, and efficient infrastructure. They used 400,000 samples for supervised fine-tuning, employed custom Actor-Critic (VAPO) and Policy Gradient (DAPO) frameworks to address instability in reinforcement learning training, innovatively used "seed verifiers" and "seed thinking verifiers" to evaluate model output quality, and achieved training efficiency improvements through the HybridFlow framework and Streaming Deployment System (SRS), reportedly increasing reinforcement learning cycle speed by 3 times.
Future Development and Industry Impact
While Seed-Thinking-v1.5 is currently unavailable for download or use, and its licensing terms are unpublished, its emergence undoubtedly intensifies competition in the reasoning AI field, setting a new standard for powerful and efficient large language models. This project is a collaborative effort of ByteDance's Seed LLM system team, led by Yong Hui Wu, with Haibin Lin as the public representative. The team plans to continue refining reinforcement learning techniques and publicly release internal benchmarks like BeyondAIME to foster broader progress in reasoning AI research.
