OpenAI has released GPT OSS, a new open-weight model family with two models: gpt-oss-120b and gpt-oss-20b. The release marks a significant departure from OpenAI's previous approach, licensing models under the permissive Apache 2.0 license.

Model Family

  • gpt-oss-120b: 117B total parameters, 5.1B active parameters - Model Card
  • gpt-oss-20b: 21B total parameters, 3.6B active parameters - Model Card

Both models use MoE architecture with 4-bit quantization (mxfp4 format). The 120B model fits on a single H100 GPU, while the 20B model runs within 16GB of memory.

Licensing: The Key Distinction

The Apache 2.0 license represents a fundamental shift from OpenAI's previous proprietary approach:

"We aim for our tools to be used safely, responsibly, and democratically, while maximizing your control over how you use them."

This contrasts sharply with restrictive licensing that limited commercial use and required proprietary APIs.

Technical Specifications

  • Architecture: Token-choice MoE with SwiGLU activations
  • Context window: 128K tokens with RoPE attention
  • Tokenizer: Same as GPT-4o
  • Capabilities: Reasoning, instruction following, tool use support
  • Quantization: MXFP4 on MoE weights only

OpenAI provides inference implementations across transformers, vLLM, llama.cpp, and ollama, including AMD ROCm hardware support.

Developer Impact

The 20B model running on consumer-grade hardware enables:

  • Local deployments without cloud dependencies
  • On-device applications
  • Private data processing
  • Reduced inference costs

The 120B model provides enterprise-grade capability while maintaining hardware efficiency.

Industry Context

OpenAI's entry with its established reputation and GPT-4o tokenizer compatibility represents a notable industry development. The announcement comes amid growing demand for accessible, well-documented models without restrictive licensing.

Availability

Models are available through:

  • Hugging Face model hub
  • Azure AI Model Catalog
  • Local inference via llama.cpp and vLLM
  • Fine-tuning capabilities
  • Partners including Dell and inference providers

Models support text-only and agentic workflows with tool use integration comparable to proprietary offerings.

Sources:

  • Hugging Face Blog: "Welcome GPT OSS, the new open-source model family from OpenAI!"
  • Model cards: gpt-oss-20b, gpt-oss-120b on Hugging Face
  • Azure AI Model Catalog integration

Verification Status: Grade A — Multiple primary sources confirm release details, licensing, and technical specifications.