Overview

gpt-oss-safeguard comes in 20B and 120B sizes and was built with ROOST (Robust Open Online Safety Tools). Instead of fixed categories, the model reads your policy at inference time, reasons about it and returns a decision with its reasoning, so policies can change without retraining.

Highlights

  • 20B (3.6B active) and 120B (5.1B active) sizes
  • Bring-your-own-policy classification
  • Full access to the reasoning behind decisions
  • The 20B model fits in 16GB of VRAM
  • Apache 2.0 license

Typical uses

Content moderation, LLM input and output filtering, and trust-and-safety labeling.

Things to consider

  • Must be used with OpenAI’s harmony response format.
  • Intended for safety tasks, not as a general chat model.
  • Download sizes refer to the default quantized versions in the Ollama library; full-precision weights are larger.

Specifications are taken from the model card on Hugging Face, the Ollama library and the provider’s announcement. Benchmark results depend on evaluation settings; see how AI benchmarks work.