GPT-OSS

Name: GPT-OSS Model Specifications
Creator: OpenAI
License: https://opensource.org/licenses/Apache-2.0

Apache 2.0 Open Source Reasoning Model

gpt-oss-120b and gpt-oss-20b are open-weight language models released by OpenAI, featuring Mixture-of-Experts (MoE) architecture. They excel at reasoning tasks with powerful tool usage and code generation capabilities, optimized for efficient deployment.

GitHub Repository Hugging Face Model Scope

GPT-OSS-120B

Large-scale reasoning model, approaching o4-mini performance

117B Total Parameters，5.1B Active Parameters
Runs on single 80GB GPU
TauBench Tool Usage:68%
Codeforces Programming:2622 Elo

GPT-OSS-20B

Lightweight edge model for local deployment

21B Total Parameters，3.6B Active Parameters
Requires only 16GB memory
Outperforms o3-mini in math competitions
Suitable for edge devices and privacy scenarios

Core Technical Features

GPT-OSS adopts advanced Mixture-of-Experts architecture and quantization technologies, significantly reducing deployment costs while maintaining high performance

Mixture-of-Experts (MoE)

120B model contains 128 experts, 20B model contains 32 experts. Each token activates only top-4 experts, dramatically reducing inference costs while preserving model capabilities.

MXFP4 Quantization Technology

MoE weights quantized to 4.25 bits, comprising 90%+ of total parameters. Combined with BF16 precision, achieving perfect balance between performance and efficiency.

Apache 2.0 License

Completely free open-source license, allowing commercial use, modification, and redistribution without license restrictions.

Harmony Chat Format

Supports multi-channel output including analysis, commentary, and final answers, enabling complex AI Agent interactions with tool calling and reasoning chain display.

Variable-length Reasoning

Supports low, medium, high reasoning levels, automatically adjusting reasoning depth based on task complexity for optimal performance-cost balance.

Powerful Tool Usage

Native support for web browsing, Python code execution, function calling, and other tools, with complete capabilities for building complex AI Agent applications.

Performance Benchmarks

Excellent performance across multiple authoritative evaluations, GPT-OSS-120B approaches OpenAI o4-mini performance

Math Competition

AIME 2024/2025

98.7%

120B model score

Exceeds o4-mini

Tool Usage

TauBench

67.8%

120B model score

Approaches o4-mini

Programming Competition

Codeforces

2622

Elo Rating

Professional level

Software Engineering

SWE-Bench

62.4%

120B model score

Real problem solving

Comparison with OpenAI Official Models

GPT-OSS-120B approaches or exceeds OpenAI o4-mini in multiple tests

OpenAI o3-mini

Baseline comparison model

GPT-OSS-120B

Open source implementation, approaching performance

OpenAI o4-mini

Target performance level

Technical Specifications

Deep dive into GPT-OSS architecture design, training process, and deployment requirements

Architecture & Training

Architecture TypeTransformer MoE

Attention MechanismGrouped Query Attention (GQA)

Context Length131,072 tokens

Tokenizero200k_harmony

Training DataTrillions of tokens

Knowledge CutoffJune 2024

Training Time2.1M H100 hours

Response FormatHarmony

Deployment Requirements

GPT-OSS-120B

80GB GPU / 60.8GB Storage

GPT-OSS-20B

16GB Memory / 12.8GB Storage

Inference FrameworkvLLM, TensorRT-LLM

Recommended Configtemperature=1.0

LicenseApache 2.0

Use Cases

GPT-OSS is suitable for various AI application development and deployment scenarios, from enterprise applications to edge computing

Code Generation & Debugging

Achieves 2622 Elo rating in Codeforces programming competitions, suitable for code generation, debugging, refactoring, and other development tasks. Supports multiple programming languages and complex algorithm implementations.

PythonAlgorithmDebug

Smart Agent Development

Scores 67.8% on TauBench tool usage evaluation, supports function calling, web browsing, Python execution, suitable for building complex AI Agent applications.

Tool CallingWeb BrowsingAutomation

Enterprise Deployment

Apache 2.0 license supports commercial use, 20B model can run on edge devices, suitable for enterprise applications with high data privacy requirements.

Data PrivacyLocal DeployCommercial

Scientific Research & Education

Scores 98.7% on AIME math competitions, excellent performance on health-related queries, suitable for research, education, and professional consulting scenarios.

Math ReasoningHealthEducation

Edge Computing & IoT

GPT-OSS-20B requires only 16GB memory, supports running on edge devices, suitable for IoT, mobile devices, and resource-constrained AI applications.

Edge ComputingIoTLow Resource

Open Source Ecosystem Contribution

Fully open-source model weights and code, supporting researchers in model improvements, fine-tuning, and innovation, advancing the open-source AI ecosystem.

Fine-tuningResearchCommunity

Safety & Compliance

GPT-OSS has undergone comprehensive safety evaluation, compliant with OpenAI Preparedness Framework standards

Safety Assessment Passed

Reviewed by OpenAI Safety Advisory Group (SAG), confirmed that the model does not reach high-risk thresholds in risk categories including biochemistry, cybersecurity, and AI self-improvement.

Instruction Hierarchy System

Implements System > Developer > User > Assistant > Tool instruction priority, ensuring the model operates according to preset security policies.

Adversarial Training

Uses deliberative alignment technology, training the model to refuse inappropriate content, resist jailbreak attacks, and follow instruction hierarchies.

Frequently Asked Questions

Common questions about GPT-OSS models and their usage

What is GPT-OSS?

GPT-OSS is OpenAI's first open-weight reasoning model series featuring 120B and 20B parameter variants with Mixture-of-Experts architecture.

Is GPT-OSS completely free?

Yes! GPT-OSS is released under Apache 2.0 license, making it completely free for commercial and non-commercial use.

How does GPT-OSS compare to other AI models?

GPT-OSS-120B approaches OpenAI o4-mini performance while being completely open-source, scoring 98.7% on AIME math competitions and 67.8% on TauBench tool usage.

What are the system requirements?

GPT-OSS-120B requires an 80GB GPU and 60.8GB storage, while GPT-OSS-20B needs only 16GB memory and 12.8GB storage.

Where can I download GPT-OSS models?

You can download GPT-OSS from GitHub, Hugging Face, or ModelScope. All platforms provide the same Apache 2.0 licensed models.

What applications support GPT-OSS?

GPT-OSS works with popular inference frameworks like vLLM and TensorRT-LLM, supporting code generation, tool usage, reasoning, and enterprise applications.

Get Started with GPT-OSS

Experience the powerful capabilities of OpenAI's first open-weight model, completely free under Apache 2.0 license

View Documentation and Code Download Model Now

117B

Parameter Scale

Apache 2.0

Open Source License

16GB+

Minimum Memory Requirement