Model reference — Documents
Text moderation checks messages, comments, reviews and uploaded copy against your own policy and flags what breaks it — abuse, threats, sexual content, self-harm, fraud, hate speech — before it reaches other people.
A moderation model produces a category and a severity, not a simple yes or no, because the action taken differs: some content is blocked outright, some is held for review, some is published with a warning. Two things decide whether a deployment works. The first is that your policy is written down precisely enough for a model to apply. The second is the balance between missing violations and blocking innocent posts — a threshold that is a business decision, not a technical one, and one we set with you and then measure.
Models in this group take text as input: a message, comment, review or document checked against your policy. Each specification table gives three hardware tiers — Minimum, the smallest setup on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is the number of sample inputs processed per hour on that hardware. Use these rates for initial sizing. Before production, benchmark your own data to validate accuracy, latency, throughput and cost. The sample input here is one message of about 100 words checked against a written policy.
Vendor: Meta
What it does: checks text against a written safety policy and returns which category it breaks and how severely. Built for moderation specifically, so it can be pointed at your own policy wording rather than a fixed list of categories.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | H100 80 GB |
| VRAM | 16 GB | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 32 GB | 48 GB | 128 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 1× H100 SXM 80 GB |
| Rate (messages/hour) | ≈ 3,000 | ≈ 9,000 | ≈ 30,000 |
Vendor: Google
What it does: the same policy-driven checking in a smaller model, tuned to be cautious. Suited to a first-pass filter running on every message before a larger model looks at what it flags.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3090 | RTX 4090 | H100 80 GB |
| VRAM | 16 GB | 24 GB | 80 GB |
| vCPUs | 8 | 12 | 24 |
| RAM | 32 GB | 48 GB | 128 GB |
| Server | 1× RTX 3090 24 GB | 1× RTX 4090 24 GB | 1× H100 SXM 80 GB |
| Rate (messages/hour) | ≈ 4,000 | ≈ 12,000 | ≈ 36,000 |
Vendor: Unitary
What it does: scores text for toxicity, insult, threat and identity attack. Very small and very fast, which makes it the practical choice for filtering high-volume comment streams in real time.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (messages/hour) | ≈ 40,000 | ≈ 160,000 | ≈ 450,000 |
Vendor: Microsoft
What it does: fitted to your own policy categories from examples your moderators have already actioned, so it enforces your decisions rather than a general standard. The most accurate option where moderation history exists.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 6 GB | 24 GB | 80 GB |
| vCPUs | 4 | 8 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (messages/hour) | ≈ 30,000 | ≈ 120,000 | ≈ 340,000 |
Vendor: Meta
What it does: applies the same policy across a hundred languages with one model, which prevents a policy being enforced strictly in English and loosely everywhere else.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 3060 | RTX 4090 | A100 80 GB |
| VRAM | 8 GB | 24 GB | 80 GB |
| vCPUs | 6 | 12 | 24 |
| RAM | 16 GB | 32 GB | 64 GB |
| Server | 1× RTX 3060 12 GB | 1× RTX 4090 24 GB | 1× A100 SXM 80 GB |
| Rate (messages/hour) | ≈ 26,000 | ≈ 100,000 | ≈ 300,000 |
Vendor: Mistral AI
What it does: judges long or ambiguous text — a three-paragraph post, a sarcastic review — and explains which sentence broke the policy, which is what a moderator needs in order to uphold or overturn the decision.
| Requirement | Minimum | Medium | High |
|---|---|---|---|
| GPU type | RTX 4090 | L40S 48 GB | H100 80 GB |
| VRAM | 24 GB | 48 GB | 80 GB |
| vCPUs | 12 | 16 | 32 |
| RAM | 48 GB | 64 GB | 128 GB |
| Server | 1× RTX 4090 24 GB | 1× L40S 48 GB | 1× H100 SXM 80 GB |
| Rate (messages/hour) | ≈ 1,300 | ≈ 3,800 | ≈ 11,000 |
Moderation is judged on two error rates at once, and the right balance depends on your audience and your legal exposure. Our consultants review your policy and a sample of your traffic, then recommend a model, the thresholds for block, hold and allow, and the volume of human review the result implies.
At the start of a project we may run a short proof of concept on a sample of your own data, measuring the accuracy and the throughput the model actually achieves on your material. That replaces the estimates on this page with real figures, so the cost and the schedule for the full engagement are known before it is committed.
Share a representative sample, expected volume, latency target and deployment location for Content moderation - text. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster.