OpenGuardrails

license

Apache-2.0(OSI) for both platform and model

source

public

self-host

yes(on-prem/private)

model

OpenGuardrails-Text-2510 open weights Apache-2.0

core-gated

ungated

Fully open source: both the guardrails platform and the safety model are Apache-2.0 and self-hostable, including on-prem deployment. The repository has since become the monorepo for the OpenGuardrails (OGR) specification and its reference implementations, carrying the Python and JavaScript SDKs and the integrations alongside it - a shift in what the product is, though the licence, the public source and the ungated core are all unchanged.

• https://github.com/openguardrails/openguardrails recorded 2026-08-13

  Repo metadata records license spdxId - Apache-2.0 and the README closes 'License. Apache-2.0'. The monorepo holds specification/ and schema/, packages/python and packages/javascript, integrations/, benchmarks/ and examples/, and the development section builds and tests the whole thing locally. The project describes itself as foundation-governed and neutral; there is no paid tier, enterprise directory or license key on the page.

• https://huggingface.co/openguardrails/OpenGuardrails-Text-2510 recorded 2026-08-13

  Model repo is ungated and carries License - apache-2.0 with a LICENSE file and two safetensors shards; cardData records base_model - Qwen/Qwen3-14B and quantization - GPTQ-4bit, which is the 3.3B quantized safety model the score records.

An early-stage project with no download figure to go on, so the level rests on GitHub stars: 24 stargazers, which bands at under 1K stars, level 1. A stars proxy never rises above 3, because a star is not a use. The project does publish SDKs to PyPI and npm, so a download route may well exist, but PyPI's statistics service was rate-limited and returned no figure.

• https://github.com/openguardrails/openguardrails recorded 2026-08-13

  stargazerCount 24 for openguardrails/openguardrails; no download, install or customer figure appears on the page.

Unusually complete for its age: it ships both an open safety model and an on-prem platform, with broad multilingual coverage. The paper supports every part of that description, but the repository has since repositioned itself as a vendor-neutral protocol and benchmark rather than a detector - "we do not build detection capability" - so the platform the paper describes and the thing the repository now ships are drifting apart. The band rests on the published model and the paper.

• https://arxiv.org/abs/2510.19169 recorded 2026-08-13

  Abstract (v2, revised 29 Oct 2025) presents OpenGuardrails as 'the first fully open-source platform that unifies large-model-based safety detection, manipulation defense, and deployable guardrail infrastructure', covering content-safety violations, model-manipulation attacks including prompt injection, jailbreaks and code-interpreter abuse, and data leakage. It records Configurable Policy Adaptation, a unified LLM-based guard architecture, a 14B base compressed to 3.3B via GPTQ, support for 119 languages, deployment as a secure gateway or API service, and Apache 2.0 for all models, datasets and deployment scripts.