Seed-OSS-36B-Instruct is a 36-billion parameter large language model developed by the Seed Team at ByteDance. It is part of the Seed-OSS series, designed for intensive long-context processing, reasoning, and agentic workflows. The model is released under the Apache 2.0 license, facilitating both research and commercial applications.
The model utilizes a decoder-only Transformer architecture with 64 hidden layers, Grouped Query Attention (GQA), and Rotary Position Embedding (RoPE). It was trained on 12 trillion tokens and provides native support for a context window of up to 512,000 tokens. This extensive context length allows the model to process large documents and maintain coherence across extended dialogues without requiring external retrieval mechanisms.
Key Features
A distinctive capability of Seed-OSS-36B-Instruct is its flexible thinking budget system. This feature allows users to dynamically adjust the reasoning length of the model's internal chain-of-thought process, balancing inference speed against the depth of analytical reasoning for specific tasks. The model is optimized for "agentic intelligence," showing proficiency in tool-use, code generation, and complex problem-solving. It also supports multilingual capabilities for international deployment.