Use cases › Deciding
Moderate and guard content
Guard models that classify prompts or responses for policy violations, jailbreaks and unsafe content.
Benchmarks that represent this job 8
A model's row on one of these third-party boards is what earns the measured label; the same benchmark name in the model's own card earns author-reported.
ToxicChatOpenAI ModerationWildGuardTestXSTestHarmBenchBeaverTailsAegisJailbreakBench
Models 13
Narrow by what you can run and pay for. Rows are ordered measured → author-reported → structural, then by watch-list score; click a column to re-sort. Every chip links to its source.
| Model | Evidence | Params | Context | Price from | Runs on | License | Providers |
|---|---|---|---|---|---|---|---|
| metask-jev-4b-policy-mix wayfind | author-reported Author reports aegis2 83.6% | 4.54B | 256K | — | Single GPU | Apache 2.0 | 0 |
| lavoir moganai | author-reported Author reports Jailbreak (toxic-chat) .825tag: guardrails | 422M | — | — | Laptop / edge | CC BY-NC 4.0 non-commercial | 0 |
| Llama-Guard-3-8B meta-llama | structural decision-model estimate p=0.99name: Guard | 8.03B | — | — | Single GPU | Llama 3.1 Community Licen… attribution | 2 |
| Qwen3Guard-Gen-0.6B Qwen | structural decision-model estimate p=1.00name: Guard | 752M | 32K | — | Laptop / edge | Apache 2.0 | 1 |
| Llama-Guard-4-12B meta-llama | structural decision-model estimate p=0.98name: Guard +1 | 12.0B | 160K | $0.14 | Single high-end GPU | Llama Community License (… attribution | 2 |
| gpt-oss-safeguard-20b openai | structural decision-model estimate p=1.00name: guard | 21.5B | 128K | $0.07 | Single high-end GPU | Apache 2.0 | 2 |
| Qwen3Guard-Gen-8B Qwen | structural decision-model estimate p=1.00name: Guard | 8.19B | 32K | — | Single GPU | Apache 2.0 | 1 |
| Qwen3Guard-Gen-4B Qwen | structural decision-model estimate p=1.00name: Guard | 4.41B | 32K | — | Single GPU | Apache 2.0 | 1 |
| Llama-Guard-3-1B meta-llama | structural decision-model estimate p=0.97name: Guard | 1.50B | — | — | Laptop / edge | Llama 3.2 Community Licen… attribution | 0 |
| Meta-Llama-Guard-2-8B meta-llama | structural decision-model estimate p=0.97name: Guard | 8.03B | — | — | Single GPU | Llama 3 Community License attribution | 1 |
| laya convaiinnovations | structural tag: moderation | 421M | — | — | Laptop / edge | Apache 2.0 | 0 |
| laya-multilingual convaiinnovations | structural tag: moderation | 322M | — | — | Laptop / edge | Apache 2.0 | 0 |
| lev interfaze-ai | structural tag: moderation | — | — | — | — | Apache 2.0 | 0 |
No model matches every constraint. Loosen one, or switch the evidence filter back to any.
How to read this
- measured a third-party leaderboard row for this model. Settings (agent harness, edit format, prompt) differ between boards; compare within a benchmark, not across.
- author-reported a number the model author published in the card or eval files. Not reproduced by us.
- structural a deterministic signal from the name, config or chat template, or a decision-model estimate. A weak signal: it says the model was built or tagged for the job, not how well it does it. "decision-model estimate" chips are a probability from the Jev decision model, not a measurement.
- Runs on is the memory class of the BF16 weights (method); choosing a class shows that class and smaller. Quantized builds may fit a class lower.
- Price from is the cheapest observed hosted input price, USD per million tokens; models with no observed price are hidden when a ceiling is set. License friction: permissive = none or attribution-only; high = non-commercial.
- Ordering within a tier is the watch-list score (Hub likes, downloads, recency), not a quality verdict.
Subscribe
Follow "Moderate and guard content"
- /feeds/for/safety-guard.xml(RSS 2.0) · one item per model with measured or author-reported evidence for this job
- /feeds/for/safety-guard.json(JSON Feed 1.1) · same items, JSON Feed
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.