Seed-OSS is a family of open-weight large language models released by ByteDance's Seed team in August 2025.1 The release centered on Seed-OSS-36B, a dense decoder-only Transformer with a developer-reported 36 billion parameters, published under the Apache 2.0 license. It marked one of ByteDance's most prominent open-weight releases, complementing the company's proprietary Doubao assistant models and its Seedream image-generation line.
Design and training
Unlike many contemporaneous Chinese open-weight releases such as Qwen3, GLM-4.5, and Kimi K2, which used mixture-of-experts designs, Seed-OSS-36B is a dense model. The Seed team reported pretraining on roughly 12 trillion tokens and a native context window of 512,000 tokens, positioning long-context handling as a headline capability.2 A distinctive feature of the Instruct variant is a configurable "thinking budget": users can cap the number of reasoning tokens the model spends before answering, allowing explicit trade-offs between latency and reasoning depth, an approach related to broader work on test-time compute in reasoning models such as DeepSeek-R1.
Variants and access
The August 2025 release comprised three checkpoints: two base models, one pretrained with synthetic instruction data and one without (the latter intended as a cleaner starting point for research on instruction tuning and post-training), plus Seed-OSS-36B-Instruct, the post-trained assistant model.2 Weights were distributed through Hugging Face and GitHub. Developer-reported benchmark results placed the Instruct model competitively among open-weight models of comparable scale on reasoning, mathematics, and coding evaluations, though these figures were self-reported and, as is common for such releases, not independently audited at launch.
Context
Seed-OSS arrived during a crowded 2025 season of open-weight releases from Chinese AI labs and Western developers alike, including OpenAI's gpt-oss. Commentators viewed the release as a signal that ByteDance, previously focused on proprietary consumer products, intended to compete directly in the open-weight ecosystem.3 The unusually long native context and the thinking-budget mechanism were the features most frequently highlighted in early coverage.