Seed-OSS-36B-Instruct
Description
Seed-OSS-36B-Instruct is a 36-billion parameter open-source large language model developed by ByteDance's Seed Team. It is designed for powerful long-context, reasoning, agent and general capabilities, and versatile developer-friendly features. The model features flexible control of thinking budget, enhanced reasoning capability, agentic intelligence, and native long context support up to 512K tokens.
This model is ready for commercial use.
Third-Party Community Consideration:
This model is not owned or developed by NVIDIA. This model has been developed and built to a third-party's requirements for this application and use case; see link to Non-NVIDIA Seed-OSS-36B-Instruct Model Card.
License and Terms of Use:
GOVERNING TERMS: This trial service is governed by the NVIDIA API Trial Terms of Service. Use of this model is governed by the NVIDIA Community Model License. Additional Information: Apache 2.0.
Deployment Geography: Global
Release Date:
Build.NVIDIA.com [09/05/2025] via link
Huggingface 08/20/2025 via link
Reference(s):
Model Architecture:
Architecture Type: Causal language model with RoPE
Network Architecture: Transformer-based decoder-only
Total Parameters: 36B
Active Parameters: 36B
Vocabulary Size: 155K
Base Model: Seed-OSS-36B-Base
Input:
Input Types: Text
Input Formats: Natural language prompts, conversational messages
Input Parameters: [One-Dimensional (1D)]
Other Input Properties: Max Input Tokens: 512K, Support for thinking budget control, tool calling, long context up to 512K tokens
Input Context Length (ISL): 512K tokens
Output:
Output Types: Text
Output Format: Natural language responses, structured tool calls
Output Parameters: [One-Dimensional (1D)]
Other Output Properties: Max Input Tokens: 512K, Chain-of-thought reasoning, thinking budget reflection, direct responses
Output Context Length (OSL): Configurable based on remaining context
Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
Software Integration:
Runtime Engines: Transformers, vLLM (>=0.10.0)
Supported Hardware:
- NVIDIA Ampere
- NVIDIA Ada Lovelace
- NVIDIA Hopper
- NVIDIA Blackwell
Preferred/Supported Operating System: Linux
Model Version(s)
v1.0
Training, Testing, and Evaluation Datasets:
Data Modality: Text
Text Training Data Size: ~12 Trillion Tokens
Training Dataset
Training Data Collection: [Hybrid: Automated, Human]
Training Labeling: [Hybrid: Automated, Human]
Training Properties: Pre-trained over 12 trillion tokens, knowledge cutoff of 07/2024, data from multiple sources including publicly available internet data, purchased data through vendor partnerships, and in-house generated data
Testing Dataset
Testing Data Collection: [Hybrid: Automated, Human]
Testing Labeling: [Hybrid: Automated, Human]
Testing Properties: Regular safety testing and adversarial testing conducted to identify and address safety vulnerabilities
Evaluation Dataset
Evaluation Benchmark Score: MMLU-Pro: 82.7, MMLU: 87.4, GPQA-D: 80.7, BBH: 89.1, AGIEval-en: 75.8, GSM8K: 93.1, MATH: 84.2, MBPP: 82.6, HumanEval: 78.8, RULER(128K): 94.6, AIR-Bench: 75.6
Evaluation Data Collection: [Hybrid: Automated, Human]
Evaluation Labeling: [Hybrid: Automated, Human]
Evaluation Properties: Safety evaluation including training data filtering, safety fine-tuning evaluation, and content safety measures assessment
Inference
Acceleration Engine: Transformers, vLLM
Test Hardware: H100
Additional Details
Key features include:
- Flexible Control of Thinking Budget: Users can adjust reasoning length dynamically
- Enhanced Reasoning Capability: Optimized for reasoning tasks while maintaining general capabilities
- Agentic Intelligence: Exceptional performance in agentic tasks like tool-using and issue resolving
- Native Long Context: Trained with up-to-512K long context natively
- Research-Friendly: Released both with and without synthetic instruction data
- Content Safety Measures: Training data filtering, safety fine-tuning, and regular safety evaluation
- International (i18n) Model: Primarily optimized for English with limited performance in other languages
- Knowledge Cutoff: July 2024
Ethical Considerations
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here
