This model was retired on 30 October 2026 and no longer accepts API requests. Recommended replacement: DeepSeek V4 Flash 0731.
GPT-OSS-20B
Retired
Powerful chain-of-thought reasoning in an efficient, highly scalable 20B parameter model.
Total Parameters
20B
Context Window
128K
Tokens
Infrastructure
Single
B200 GPU
License
Apache
2.0
Scalable Open Reasoning
GPT-OSS-20B provides powerful chain-of-thought reasoning in an incredibly efficient 20B parameter model. Designed for single-GPU deployment while maintaining sophisticated reasoning capabilities, this Apache 2.0 licensed model offers the perfect balance of performance and resource efficiency. It utilizes a compact Mixture-of-Experts (MoE) design with SwiGLU activations, native FP4 quantization for optimal inference speed, and an adjustable reasoning effort level for task-specific optimization.
Compact MoE — Single-GPU Ready
Built for efficient reasoning
Development Applications
Accelerate engineering with rapid prototyping, robust code generation, automated API design, and system integration testing.
Business Solutions
Streamline enterprise operations through customer support automation, advanced content generation, and sophisticated market research analysis.
Educational Use Cases
Deploy interactive tutoring systems, support curriculum development, and assist with academic writing and research methodology guidance.
Edge & Scalable Deployment
Leverage cost-effective single-GPU operations for reduced infrastructure requirements, making it ideal for edge computing and distributed processing.
Efficient Intelligence
Proven performance across reasoning, coding, and agentic workflows for the 20B weight class.
Intelligence Index
Better than 58% of models
GPQA Diamond
Better than 64% of models
τ²-Bench Telecom
Better than 50% of models
| Category | Benchmark | Score | Description |
|---|---|---|---|
| Reasoning | GPQA Diamond | 73% | Graduate-level scientific reasoning |
| Reasoning | Humanity's Last Exam | 10% | Humanity's Last Exam |
| Reasoning | τ²-Bench Telecom | 44% | AI agents in dual-control scenarios |
| Reasoning | AA-LCR | 31% | Long context reasoning evaluation |
| Reasoning | IFBench | 47% | Instruction-following accuracy |
| Coding | SciCode | 28% | Python for scientific computing |
| Coding | Terminal-Bench Hard | 11% | Agentic coding & terminal use |
| Knowledge | AA-Omniscience Accuracy | 16% | Proportion of correctly answered questions |
Metrics sourced from Artificial Analysis.