Skip to main content
Raven is PolyAI’s proprietary LLM, built for real-time customer conversations. Sub-300ms latency across 24+ languages. Raven runs the majority of PolyAI deployments. Select it in Voice configuration or Chat configuration, or compare it with other models on the Model page.

​
Why Raven

General-purpose models (GPT, Claude) are trained for broad text tasks. They can handle customer service, but require heavy prompting to behave reliably and add latency a voice channel cannot absorb. Raven is built specifically for customer conversations:
  • Sub-300ms latency. Fast enough for live voice with consistent response times.
  • Trained on phone conversations. Handles interruptions, mispronunciations, and partial utterances.
  • Conversational behavior built in. Staying on topic, grounding answers in your knowledge, and asking for clarification when unsure do not require prompting.

Conversation-native

Built for customer service across voice and chat. Add raw information to your knowledge – Raven converts it into natural conversational responses without extra prompting.

Faster

Sub-300ms median latency. Consistent response times – no long-tail spikes.

More accurate

Higher accuracy on PolyAI’s customer service benchmarks. Fewer errors in tool calling and knowledge grounding.

Multilingual

24+ languages with near-perfect language consistency. Set the response language – Raven speaks it, even with English-only prompts.

​
Additional capabilities

Date and time logic – handles relative dates, scheduling, and format conversions that trip up general-purpose models. Reliable tool calling – trained on real Agent Studio projects. Calls functions with correct parameters; doesn’t confuse responding with acting. No hallucination – grounded in your knowledge. Says “I don’t know” rather than inventing answers. Agent Studio native – understands topics, flows, and PolyAI’s knowledge retrieval patterns by default.

​
Raven 3.5

Latest Raven model. Supports voice and chat. Recommended for all new deployments.
  • Auto-reasoning – automatically decides when to think deeper before responding, improving accuracy on complex tasks like date calculations without adding latency on simple turns
  • Out-of-domain detection – identifies when a request falls outside the agent’s scope, enabling cleaner handoffs and knowledge gap tracking
  • Built-in safety – guardrails against misuse, with built-in protection against hallucinations
  • Custom style following – respects custom persona and style instructions, including emotion tags for TTS, formatting rules, and channel-specific tone
  • 24+ languages – more natural multilingual outputs than earlier Raven versions, with near-perfect language consistency
Raven V3 is deprecated. Older Raven versions now route to Raven 3.5 automatically. Existing deployments keep working, but you should select Raven 3.5 directly in your agent configuration.

​
Supported languages

Raven supports the following languages:
Arabic, Bulgarian, Cantonese, Croatian, Czech, Dutch, English, French, German, Greek, Hindi, Hindi (Romanized/Hinglish), Italian, Japanese, Korean, Mandarin (China), Mandarin (Taiwan), Polish, Portuguese (Brazil), Portuguese (Portugal), Serbian, Spanish (US), Swedish, Turkish
You can keep all your prompts and knowledge in English and set the response language to your target language. Raven responds consistently in the target language. Quality improves further if you translate prompts and add examples in the target language.

​
Getting started

Select Raven 3.5 in Voice configuration or Chat configuration.

Model selection

Compare Raven with OpenAI and Amazon Bedrock models.

Training data

Transparency on datasets used to train Raven.

Bring your own model

Connect your own LLM endpoint to PolyAI.
Last modified on July 10, 2026