Seed-Thinking-v1.5

What it is

Seed-Thinking-v1.5 is a reasoning model from ByteDance Seed. It is a mixture-of-experts (MoE) model, meaning that only part of the network — 20B of its 200B total parameters — is activated for a given input ^[1][2]. Like other reasoning models, it was trained with reinforcement learning to think before answering, producing intermediate reasoning before the final response ^[1][2]. The public GitHub repository carries the technical report and the benchmark tables, but lists no model weights ^[2].

Where it came from

The technical report is titled "Seed1.5-Thinking: Advancing Superb Reasoning Models with Reinforcement Learning". It was submitted to arXiv on 10 April 2025 and revised on 22 and 29 April. Jiaze Chen leads an author list of more than 270 people ^[1].

Reported results

The headline scores given for the model are 86.7 on AIME 2024, 74.0 on AIME 2025, 77.3 on GPQA Diamond and 55.0 pass@8 on Codeforces ^[2]. The report places these next to three competitors: DeepSeek R1 (79.8 on AIME 2024, 71.5 on GPQA Diamond), OpenAI o3-mini (87.3 on AIME 2024) and Gemini 2.5 Pro (92.0 on AIME 2024) ^[2]. Read against those numbers, Seed-Thinking-v1.5 sits above DeepSeek R1 on both AIME 2024 and GPQA Diamond, and below o3-mini and Gemini 2.5 Pro on AIME 2024 ^[2]. The repository's comparison table gives the model 74.0% on AIME 2025, against 65.0% for DeepSeek R1 ^[2].

Not every comparison goes its way. On Codeforces pass@8, the repository table puts OpenAI o3-mini at 67.5%, above the 55.0% listed for Seed-Thinking-v1.5 ^[2].

The report also gives a set of lower scores on harder or more open-ended evaluations: 47.0 on SWE-bench Verified, 39.9 on ARC-AGI and 12.9 on SimpleQA ^[2].

The model was not only measured on reasoning tasks. On non-reasoning tasks it beat DeepSeek R1 by 8% in win rate ^[1].

Building harder tests

AIME 2024 no longer separated the strongest models well, so the team built a harder internal evaluation set called BeyondAIME, on which the model scored 48.0, and a Codeforces-based benchmark; both were planned for public release ^[1][2]. BeyondAIME did distinguish the models more sharply, but not in this model's favour: on the repository table OpenAI o3-mini reaches 63.6% against 48.0% for Seed-Thinking-v1.5 ^[2].

Where it sits

Seed-Thinking-v1.5 is a snapshot of how frontier reasoning models were being built and measured around the time of its report: a sparse mixture-of-experts backbone, reinforcement learning on top of it to produce reasoning before answering, and evaluation on competition mathematics (AIME), science questions (GPQA Diamond), programming contests (Codeforces) and software engineering (SWE-bench Verified) ^[1][2]. Its results are mixed by task — competitive on AIME and GPQA, behind OpenAI's o3-mini on Codeforces — and the accompanying discussion of benchmark saturation, together with the plan to release BeyondAIME, records a moment when the standard tests were running out of headroom ^[1][2]. Because the repository publishes reports and tables rather than weights, the model is documented for comparison rather than distributed for use ^[2].