Details on GPT-OSS - Free Open-Source LLM That Runs on Local Computers

Introduction to OpenAI's GPT-OSS Model

On August 5, 2025, OpenAI officially introduced GPT-OSS-120B and GPT-OSS-20B, two open-source language models with open weights and high performance.

Key highlights:

• Apache 2.0 License: allows free use and modification, including for commercial purposes.
  
• Superior performance compared to other open models of similar size.
  
• Optimized for inference and powerful tool use.
  
• Compatible with common hardware:
  
  • 120B runs on an 80 GB GPU.
    
  • 20B requires only 16 GB RAM.
    

Training the GPT-OSS Models

The GPT-OSS models were trained by OpenAI using advanced pre-training and post-training techniques, focusing on reasoning capabilities, performance optimization, and practical applicability across various deployment environments. These are OpenAI's first open-weight language models since GPT-2 (2019), following the release of other open models like Whisper and CLIP.

In terms of architecture, GPT-OSS employs a Mixture of Experts (MoE) approach to reduce the number of active parameters per token:

• GPT-OSS-120B: 5.1 billion active parameters/token.
  
• GPT-OSS-20B: 3.6 billion active parameters/token.
  

The training dataset primarily consists of high-quality English text, with a focus on STEM, programming, and general knowledge, encoded using the o200k_harmony token set (also used for o4-mini and GPT-4o).

The post-training phase includes supervised fine-tuning and RLHF to align with the OpenAI Model Spec, incorporating Chain-of-Thought reasoning training and tool usage.

Both models support three inference levels—low, medium, and high—allowing developers to flexibly balance speed and accuracy with a simple system prompt setting.

Performance Evaluation of GPT-OSS Models

OpenAI benchmarked GPT-OSS-120B and GPT-OSS-20B on standard academic datasets to assess programming, competitive math, healthcare, and tool usage capabilities, comparing them directly with o3, o3-mini, and o4-mini.

• GPT-OSS-120B:
  
  • Outperforms o3-mini and matches or slightly exceeds o4-mini on Codeforces, MMLU, HLE, and TauBench.
    
  • Surpasses o4-mini on HealthBench and AIME 2024 & 2025 tasks.
    
• GPT-OSS-20B:
  
  • Matches or exceeds o3-mini on the same test sets, despite its smaller size.
    
  • Particularly strong in competitive math and healthcare tasks.
    

GPT-OSS-120B approaches (and sometimes surpasses) smaller proprietary models in many scenarios, while GPT-OSS-20B offers a lightweight yet competitive performance option, ideal for local deployment and low-cost solutions.

Conclusion

The launch of GPT-OSS-120B and GPT-OSS-20B marks a significant milestone for OpenAI in bringing open-weight language models to the community. With performance rivaling proprietary models, hardware-optimized architecture, and stringent safety standards, GPT-OSS not only creates opportunities for large enterprises but also empowers individuals, startups, and small organizations to access cutting-edge AI technology.

This duo combines exceptional reasoning power, flexible customization, and cross-platform compatibility, enabling anyone to build, deploy, and optimize AI solutions on their own infrastructure. In a landscape where the demand for transparent, democratized, and accessible AI is growing, GPT-OSS stands as a testament to OpenAI’s vision: powerful AI must belong to everyone.