OpenAI gpt-oss-safeguard models released via Ollama

TL;DR

OpenAI, Ollama, and ROOST announced the open‑source gpt-oss-safeguard models (20 B and 120 B parameters) under an Apache 2.0 license, providing a reasoning‑based safety classifier that can interpret arbitrary user‑defined policies and expose its chain‑of‑thought decisions.

Model Availability and Licensing

Takeaway: The models are freely downloadable, runnable via Ollama, and carry a permissive Apache 2.0 license, removing copyleft and patent concerns for commercial or experimental use.

  • Two model sizes are offered: gpt-oss-safeguard:20b and gpt-oss-safeguard:120b.
  • Installation steps are simple: download Ollama, then run ollama run gpt-oss-safeguard:20b (or :120b).
  • Apache 2.0 licensing allows unrestricted modification, redistribution, and commercial deployment.

Core Technical Features

Takeaway: The models are purpose‑built for safety reasoning, supporting custom policies, transparent reasoning output, and adjustable inference effort.

Safety‑Focused Training

  • Trained and fine‑tuned explicitly to reason about safety classification tasks such as input‑output filtering, online content labeling, and offline trust‑and‑safety pipelines.

Bring‑Your‑Own‑Policy (BYOP)

  • The model accepts a textual policy description at inference time and applies it across diverse products without additional engineering.

Reasoned Decisions Instead of Raw Scores

  • Returns a full chain‑of‑thought (CoT) explanation, enabling developers to debug and audit decisions.
  • The raw CoT output is intended for safety practitioners, not end‑users.

Configurable Reasoning Effort

  • Users can select low, medium, or high reasoning effort, trading off latency against depth of analysis.

Evaluation Results

Takeaway: OpenAI’s internal and external benchmarks show that gpt‑oss‑safeguard accurately classifies text against multiple simultaneous policies and performs competitively on established moderation datasets.

  • Internal Multi‑Policy Test: Accuracy is measured only when the model’s label matches the gold set for all supplied policies, a stringent criterion.
  • OpenAI Moderation Dataset (2022): The models were evaluated on the same dataset released with OpenAI’s 2022 moderation research paper.
  • ToxicChat Benchmark: Performance was also measured on the public ToxicChat benchmark, which reflects real‑world user queries to open‑source chatbots.

Internal multi‑policy evaluation OpenAI moderation benchmark ToxicChat benchmark

Community and Organizational Context

Takeaway: The release is backed by ROOST, a non‑profit dedicated to open‑source safety tooling, underscoring a broader movement toward transparent, community‑driven AI safety infrastructure.

“gpt‑oss‑safeguard is the first open source reasoning model with a ‘bring your own policies and definitions of harm’ design. Organizations deserve to freely study, modify and use critical safety technologies and be able to innovate. In our testing, it was skillful at understanding different policies, explaining its reasoning, and showing nuance in applying the policies, which we believe will be beneficial to builders and safety teams.” – Vinay Rao, CTO of ROOST

About ROOST

  • ROOST (Robust Open Online Safety Tools) is a 2025‑founded non‑profit that supplies high‑quality, open‑source safety tools to digital organizations.
  • It is supported by a coalition of technology companies, philanthropic groups, and academia.

Getting Started Resources

Takeaway: Developers can follow official guides and community repositories to integrate the models into their pipelines.

Implications for the AI Safety Landscape

Takeaway: By releasing a high‑capacity, policy‑driven safety model under a permissive license, OpenAI and its partners lower the barrier for organizations to implement robust content moderation, potentially accelerating the adoption of responsible AI practices across industry and research.

  • The BYOP design reduces the need for bespoke safety model training, saving resources for smaller teams.
  • Transparent reasoning fosters trust and facilitates regulatory compliance where explainability is required.
  • Open licensing encourages community audits, extensions, and integration with diverse ecosystems, mitigating the risk of vendor lock‑in.

Sources

Related

  • Dispatch
    gpt-oss-safeguard release notes / what's new

    OpenAI has released gpt-oss-safeguard, a set of open-weight reasoning models designed for safety classification tasks that allow developers to apply custom policies at inference time.

  • Dispatch
    OpenAI gpt-oss-safeguard Technical Report

    OpenAI has released gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, open-weight reasoning models designed specifically for content classification based on provided policies.

  • Dispatch
    OpenAI gpt-oss-120b and gpt-oss-20b Release

    OpenAI has released gpt-oss-120b and gpt-oss-20b, two open-weight reasoning models under the Apache 2.0 license designed for agentic workflows and customizable reasoning effort.

  • Dispatch
    OpenAI gpt-oss Release Notes

    OpenAI has released gpt-oss-120b and gpt-oss-20b, two open-weight reasoning models licensed under Apache 2.0 that deliver frontier-level reasoning and tool-use capabilities optimized for consumer hardware.

  • Dispatch
    OpenAI gpt-oss Release on Ollama

    Ollama has partnered with OpenAI to integrate the gpt-oss open weight models, offering 20B and 120B parameter versions optimized for local execution via the MXFP4 quantization format.