ByteDance Seed publishes the Seed-Thinking-v1.5 reasoning model

Background

A reasoning model is a language model trained to work through a problem in steps before it gives an answer. Such models are usually tested on competition mathematics, programming and science questions, where the answer can be checked. AIME is a set of hard high-school mathematics competition problems that is widely used for this purpose.

What happened

On 10 April 2025 ByteDance Seed posted the technical report for its reasoning model Seed-Thinking-v1.5 to arXiv. The report is also titled Seed1.5-Thinking.^[1]

The model uses a mixture-of-experts design, with 20 billion parameters activated for each input and 200 billion in total.^[1]

Key results

The report gives 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA. Codeforces is a platform for programming contests, and GPQA is a set of difficult science questions.^[1]

It beat DeepSeek R1 by 8% in win rate on non-reasoning tasks.^[1] On AIME 2025 it scored 74.0, against 65.0 for DeepSeek R1 and 86.5 for OpenAI o3-mini.^[2]

Benchmark

AIME 2024 no longer separated the best models well, so the team built a harder internal benchmark called BeyondAIME.^[1]

In a mixture-of-experts model the network holds many expert sub-networks, and a routing step selects a few of them for each input. The total parameter count describes the stored model, while the activated count describes how much work is done for one input. Reasoning models of this kind are commonly trained with reinforcement learning on problems whose answers can be checked, which is why mathematics and programming tests are common.

Reasoning models usually produce a long chain of intermediate steps, and the length of that chain is part of what makes them costly to run. Competition-style tests are attractive for evaluation because each question has a clear answer and the same questions can be given to every model.

Significance

The report sets its AIME 2025 result directly against DeepSeek R1 and OpenAI o3-mini, so readers can see its standing on the same test.