GPT-OSS — Open-Weight Language Models by OpenAI

Introduction

GPT-OSS is OpenAI's release of open-weight language models in two sizes: a 120B parameter flagship and a 20B parameter efficient variant. These models can be downloaded, run locally, fine-tuned, and deployed without API dependency.

What GPT-OSS Does

• Provides fully open model weights for local inference
• Offers a 120B model for maximum capability and a 20B model for efficiency
• Supports text generation, instruction following, and code completion
• Enables fine-tuning on custom datasets without API costs
• Integrates with standard inference frameworks like vLLM and TGI

Architecture Overview

GPT-OSS models use a transformer architecture with grouped-query attention and rotary position embeddings. The 120B model uses a mixture of experts design for efficient inference relative to its parameter count. Weights are distributed in safetensors format compatible with Hugging Face Transformers, vLLM, and llama.cpp (via GGUF conversion).

Self-Hosting & Configuration

• Download weights from the GitHub releases or Hugging Face Hub
• The 20B model fits on a single GPU with 24 GB VRAM (quantized)
• The 120B model requires multi-GPU setup or quantization to 4-bit
• Runs with vLLM, TGI, or Hugging Face Transformers out of the box
• Tokenizer and chat template are included in the repository

Key Features

• Open weights with permissive licensing for commercial use
• Two model sizes for different hardware and latency requirements
• Compatible with OpenAI API format for drop-in replacement
• Supports GGUF quantization for CPU and edge deployment
• Fine-tuning scripts included for custom domain adaptation

Comparison with Similar Tools

• Llama 3 — Meta's open model; GPT-OSS offers OpenAI's architecture and training approach
• Mistral — efficient open models; GPT-OSS provides a larger 120B flagship option
• Qwen — Alibaba's open LLM; GPT-OSS benefits from OpenAI's training data pipeline
• DeepSeek-V3 — strong open model; GPT-OSS integrates natively with OpenAI tooling

FAQ

Q: Is GPT-OSS the same as GPT-4 or GPT-5? A: No, GPT-OSS is a separate model family designed specifically for open release with its own architecture.

Q: Can I fine-tune these models? A: Yes, fine-tuning scripts and documentation are provided in the repository.

Q: What is the license? A: The models use a permissive license that allows commercial use with attribution.

Q: How does the 20B model compare to the 120B? A: The 20B model is suitable for most tasks; the 120B excels at complex reasoning and long-context generation.

Sources

• https://github.com/openai/gpt-oss
• https://github.com/openai/gpt-oss/blob/main/README.md