Zen Guard: Multilingual AI Safety Classification

Zen Guard is a family of safety classification models built for production AI systems. The flagship is an 8B classifier covering 119 languages and 9 harm categories, with a three-tier severity system and 5ms/token latency suitable for real-time content streams.

Safety classification that only works in English is not safety classification -- it is a liability with a gap in it. Zen Guard was built to cover the full scope of deployed language.

The Zen Guard Family

Three models designed for different deployment contexts:

Model	Parameters	Languages	Use Case
Zen Guard	8B	119	Batch and async classification
Zen Guard Gen	8B	119	Generative explanations
Zen Guard Stream	4B	119	Real-time streaming safety

Zen Guard is the core classifier: binary and severity-graded outputs for API responses, dataset filtering, and human review queuing.

Zen Guard Gen produces natural-language explanations alongside classifications. When a safety system needs to surface reasoning to a human reviewer or generate policy-compliant refusal messages, Zen Guard Gen provides the generative layer.

Zen Guard Stream is 4B, optimized for 5ms/token latency. Deploy it inline in a streaming inference pipeline and classify tokens as they arrive. This enables real-time blocking without a separate moderation round-trip.

Nine Harm Categories

Zen Guard classifies across nine categories:

1. Violent content -- Graphic violence, threats, incitement
2. Sexual content -- Explicit material, non-consensual content
3. Hate speech -- Bias, discrimination, dehumanization
4. Self-harm -- Instructions or encouragement for self-injury
5. Dangerous information -- Weapons, synthesis routes, harmful instructions
6. Privacy violations -- PII exposure, doxxing, surveillance content
7. Deception -- Fraud, impersonation, manipulation
8. Illegal activity -- Content facilitating or promoting illegal acts
9. Child safety -- CSAM and child exploitation content

Each category returns an independent confidence score. Multi-label output means a single message can trigger multiple categories simultaneously.

Three-Tier Severity

Tier	Label	Recommended Action
0	Safe	Pass through
1	Controversial	Log, flag for review, or soft-filter depending on policy
2	Unsafe	Block, quarantine, escalate

The controversial tier is deliberate. Most real-world deployments need more granularity than a binary safe/unsafe gate. Tier 1 content includes political commentary, adult topics that may be contextually appropriate, and content that varies in acceptability by jurisdiction or use case. Tier 2 is unambiguous -- block it.

Streaming Integration

At 5ms/token, Zen Guard Stream runs inline without materially impacting generation latency:

import httpx

async def safe_stream(prompt: str):
    async with httpx.AsyncClient() as client:
        async with client.stream("POST", "https://api.hanzo.ai/v1/chat/completions",
            json={"model": "zen-guard-stream", "stream": True, "messages": [...]}) as r:
            async for chunk in r.aiter_text():
                safety = classify_chunk(chunk)
                if safety.tier == 2:
                    yield "[Content blocked by Zen Guard]"
                    return
                yield chunk

The 4B architecture fits on a single A10G alongside most inference workloads, making co-deployment straightforward.

Language Coverage

119 languages including full coverage for all major world languages, regional variants, and commonly used low-resource languages. Training data includes native-language harmful content -- not machine-translated examples -- which is critical for accurate classification in non-English languages where harmful phrasing often does not map directly from English idioms.

Get Zen Guard

• HuggingFace: huggingface.co/zenlm
• Hanzo Cloud API: api.hanzo.ai/v1/moderations
• Zen LM: zenlm.org -- integration guides

---

Zach Kelling is the founder of Hanzo AI, Techstars '17.