Detoxify

Detoxify is a free, self-hosted Python library for classifying toxic text. It is best suited to developers and researchers who can run models themselves and want to flag potentially harmful comments. Its model choices and multilingual coverage are useful, but profanity-linked bias makes its predictions unsuitable as unreviewed moderation decisions.

Overview

The library pairs trained models with code to classify comments, with intended uses spanning research, fine-tuning on representative datasets, and helping moderators identify harmful content faster. That makes it a focused text-classification component rather than a full moderation system: it has no custom policies or human-review workflow, and it does not handle images, audio, or video.

Models were trained for the 2018, 2019, and 2020 Jigsaw toxic-comment challenges using Wikipedia Comments and Civil Comments data. The README warns that words associated with profanity may be labeled toxic regardless of tone or intent, potentially biasing results against vulnerable minority groups. Treat scores as signals for review, not definitive judgments.

Key features

Three model variants

Detoxify offers original, unbiased, and multilingual models based on BERT, RoBERTa, and XLM-RoBERTa, respectively. Depending on the model, outputs include categories such as threats, obscenity, insults, identity attacks, and explicit sexual content. The range gives developers options for different classification needs, but category coverage varies by model; there is no custom-policy layer to tailor labels to a community's rules.

Python inference and runtime

Install the library with pip and use Python to predict labels for one string or a list of strings. Inference supports CPU by default and accepts a PyTorch device such as CUDA. It relies on Hugging Face Transformers and PyTorch Lightning, so it fits projects already able to manage that stack better than teams seeking a standalone moderation dashboard. Training additionally requires the Kaggle API to download data.

Multilingual coverage and API

The multilingual model is intended for English, French, Spanish, Italian, Portuguese, Turkish, and Russian. This is a useful reach for those languages, not evidence of broad coverage beyond them. The repository says API access can be requested by emailing [email protected], which is a separate route from running the self-hosted library.

Pricing

Open-source library: free. The plan includes trained models and code for toxic-comment classification, with API access by email. There is no free trial because the library is free. The price removes a software fee, but it does not change the practical requirements: teams need to run the Python stack themselves, and no request quota or seat allowance is part of this plan.

The project is distributed under the Apache-2.0 license. It is a strong fit when self-hosting and implementation work are acceptable; organizations needing managed moderation operations, configurable policies, or broader media coverage should look elsewhere.

Platforms

Detoxify supports Linux and self-hosted deployment. Its Python interface and local inference options suit teams integrating classification into their own systems, rather than buyers looking for a hosted web application. Text moderation is supported; image, audio, and video moderation are not.

Who it's for

Choose Detoxify if you are a developer or researcher who wants open-source toxic-text models, can work in Python, and values running inference in your own environment. The multilingual model can help with the seven intended languages, while CPU support provides a default runtime without requiring CUDA. It is a weaker fit for teams that need a turnkey queue, human escalation, custom rules, or moderation across media types.

Pros and cons

Pros

• Free, self-hosted code and models: useful for teams that want to integrate classification without a software subscription.
• Three model variants: offers choices based on BERT, RoBERTa, and XLM-RoBERTa, with toxicity-related labels that vary by model.
• Seven-language intent for the multilingual model: covers English, French, Spanish, Italian, Portuguese, Turkish, and Russian.
• CPU by default, with CUDA accepted: supports inference on CPU while allowing a PyTorch device to be supplied.

Cons

• Profanity can be a misleading signal: the project warns that tone and intent may not prevent profanity-associated words from being classified as toxic.
• Narrow moderation scope: it handles text only and provides neither custom policies nor a human-review workflow.
• Requires a Python environment and dependencies: inference uses Transformers and PyTorch Lightning; training adds the Kaggle API requirement.

Alternatives

AI Content Moderation Software is a broader directory for comparing moderation tools.

Pick Tisane API instead if you want a freemium option with API, web, desktop, and self-hosted platforms; its Community plan includes 50,000 requests per month at 10 hits per minute, but excludes technical support, CORS support, LM Direct Access, and customization.

Choose Google Cloud Vision API when image analysis is the priority; its pay-as-you-go plan bills by image and feature, unlike Detoxify's text classification focus.

WebPurify is another moderation option.

Amazon Rekognition Content Moderation may suit teams exploring a video-capable API; its Free Tier lasts 12 months from account creation and includes 60 video analysis minutes per month, including Content Moderation.

Hive Moderation is an alternative for teams seeking access to 10+ models and selected-model pay-as-you-go, with $50+ in free credits after adding a payment method.

PicPurify offers a freemium API and web option, with 2,000 free units on its FREE plan.

Sightengine is another freemium API and web alternative; its free plan allows 1 request per second and excludes video processing and live-stream support.

Consider Llama Guard if you want a free self-hosted model; its listed Llama Guard 3-8B option requires model weights.

Verdict

Detoxify is a sensible choice for developers and researchers who want free, self-hosted toxic-text classification and can treat model outputs as prompts for moderation rather than final calls. Its model variety, multilingual option, and Python integration are the main reasons to choose it. Look elsewhere if you need multimodal coverage, custom policies, a human-review workflow, or stronger assurance that context will be reflected in toxicity labels.