ByteDance Seed open-sources Seed-OSS-36B under Apache-2.0

Release

On 20 August 2025 ByteDance Seed released three models on Hugging Face: Seed-OSS-36B-Base, Seed-OSS-36B-Base-woSyn and Seed-OSS-36B-Instruct.^[1] Hugging Face is a platform where researchers and developers publish and download open models.

Size and training

The models have 36 billion parameters and were trained on 12 trillion tokens.^[1] Their native context length is up to 512K tokens, which is the amount of text a model can take in at one time.^[1]

License

The models were released under the Apache-2.0 license, which allows use, modification and redistribution with the license notice kept.^[1]

Features

Users can set a thinking budget that caps the length of the model's reasoning.^[1]

Results

The Instruct model scored 91.7 on AIME24, a set of hard mathematics competition problems, and 56 on SWE-bench Verified (OpenHands). SWE-bench Verified tests whether a model can resolve real issues from open-source software projects, and OpenHands is the agent framework used to run that test.^[1]

Base-woSyn

The Base-woSyn model was pre-trained without synthetic instruction data. It was released alongside the main base model.^[1]

A context window of this length lets a model take in a very long document or a large codebase at once. Long contexts need more memory when the model runs, which is one reason the length is reported separately from the parameter count. A thinking budget is a practical control, because it lets the person running the model decide how much reasoning to allow for each request.

A base model is trained on general text before any instruction tuning, while an Instruct model has been tuned to follow requests. Both forms are useful, because a base model is often the starting point for further training by others.

Significance

The release gave developers a large open model with a long context window under a permissive license. Downloads from Hugging Face let them run and test the models directly.