Seed1.5-Thinking
A reasoning model from ByteDance Seed using reinforcement learning
ByteDance Seed released Seed1.5-Thinking in April 2025 to improve logical deduction and coding skills. It uses a mixture of experts architecture with 200 billion total parameters but activates only 20 billion per token.
- Uses 20 billion active parameters out of 200 billion total
- Scored 86.7 on the AIME 2024 math benchmark
- Achieved 55.0 on Codeforces programming tasks
- Released BeyondAIME and Codeforces benchmarks publicly
What it is
Seed1.5-Thinking is a language model designed to pause and reason before generating an answer. The team at ByteDance Seed introduced it in April 2025. It aims to solve complex problems in science, technology, engineering, and mathematics, as well as programming tasks. The system employs a mixture of experts design, which means it has a large total parameter count but uses a smaller subset for each calculation. This approach allows it to handle difficult logical challenges while remaining efficient compared to dense models of similar total size.
How it came about
The development team published their findings in a paper submitted to arXiv in April 2025. The work represents an effort to advance reasoning capabilities through reinforcement learning techniques. The authors, led by researchers at ByteDance Seed, focused on creating a model that could generalize beyond specific test sets. They also created new internal evaluation tools to measure performance more broadly than existing standards allowed.
Inside the model
The model relies on a mixture of experts architecture. It contains 200 billion parameters in total, yet only activates 20 billion during inference. This sparse activation helps manage computational costs while maintaining high capacity. The training process emphasises reinforcement learning to refine its thinking steps. The developers claim this method improves the quality of the reasoning chain before the final output is produced.
What it can do
The model shows strong results in technical domains. It scored 86.7 on the AIME 2024 mathematics benchmark. On Codeforces, a platform for competitive programming, it achieved a score of 55.0. For the GPQA benchmark, which tests graduate-level science questions, it reached 77.3. The developers state it outperforms DeepSeek R1 by eight percent in win rate on tasks that do not strictly require deep reasoning, suggesting wider utility.
How to use it
Users can try the model via a trial link provided by the developers. The paper indicates that the model is available for testing, though specific licensing terms for weights are not detailed in the abstract. The focus is on demonstrating the capabilities through public benchmarks and shared evaluation tools.
Why it matters
The team released two new benchmarks, BeyondAIME and Codeforces, to the public. These datasets aim to help other researchers assess generalized reasoning abilities more effectively. By sharing these tools, the developers hope to support future work in evaluating how well models adapt to diverse problem types outside standard tests.
Questions people ask about Seed1.5-Thinking
What makes Seed1.5-Thinking different from other models?
It uses a mixture of experts architecture with 200 billion total parameters but only activates 20 billion per token. It focuses on reasoning through thinking steps before answering, using reinforcement learning to improve performance.
How did it perform on coding and math tests?
The model scored 86.7 on AIME 2024 for mathematics and 55.0 on Codeforces for programming. It also achieved 77.3 on the GPQA science benchmark, showing strong capabilities in STEM fields.
Who developed this model?
ByteDance Seed, a research team within ByteDance, developed the model. They published their findings in April 2025, detailing the use of reinforcement learning to enhance reasoning skills.
Good reading and sources
There is no English Wikipedia article on Seed1.5-Thinking yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2504.13914. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.