Hydra
AI model · Robotics / world model · Audio / speech → Audio / speech by Smallest.ai
Native speech-to-speech world model for real-time voice intelligence.
A native speech-to-speech world model for real-time, full-duplex voice agents. It enables natural, low-latency conversations with concurrent listening and thinking over a single WebSocket.
Model specs
The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.
- Languages
- English
Category: Conversational AI
- Modalities
- Audio → Audio
- Access
- API (WebSocket)
- Released
- 2026-04-30 (Research announcement)
- Languages
- English (at launch)
- Architecture
- Asynchronous World Model
- Communication Mode
- Full-Duplex
- Native speech-to-speech architecture
- True full-duplex conversations with natural barge-in
- Asynchronous world model on compressed latent representations
- Single WebSocket for audio-in and audio-out
Overview
Hydra is Smallest AI's native speech-to-speech model designed for real-time, full-duplex voice interactions. This architecture allows the model to listen, think, and speak concurrently, replicating human-like conversational behavior and preserving paralinguistic richness such as emotion and prosody. Built for production-grade voice agents, Hydra enables audio-in to audio-out communication over a single WebSocket to minimize latency and simplify orchestration. The model is currently in beta and is optimized for fluid conversational experiences and fast tool calling.
Conversational Intelligence
- Native Speech-to-Speech: Hydra is a native speech-to-speech model that processes audio directly without converting it to intermediate text representations.
- Full-Duplex Communication: The model supports true full-duplex conversations, allowing users to interrupt the agent naturally at any time.
- Asynchronous World Model: Hydra operates as an asynchronous world model on compressed latent representations to replicate human-like listening and thinking.
- Concurrent Processing: The architecture enables the model to listen, think, and speak concurrently rather than waiting for a user to finish.
Developer Capabilities
- Optimized Tool Calling: Hydra is designed to deliver fluid, real-time voice experiences with faster tool calls that keep conversations moving.
- Single WebSocket Pipeline: The model enables audio-in to audio-out communication over a single WebSocket without the need to assemble a complex pipeline.
- Paralinguistic Fidelity: The architecture is built to handle paralinguistic richness, preserving emotional and prosodic information.
- Production-Ready Architecture: The model is built for production-grade voice agents that require fluid, low-latency conversational intelligence.
Use cases
- Developers build full-duplex voice agents that can handle natural interruptions and barge-in during phone calls.
- Teams use Hydra to preserve emotional and prosodic information in voice interactions without the lossy conversion to text.
- Enterprises deploy real-time voice experiences that require fast tool calling and concurrent processing of audio signals.
Pricing
- Hydra Beta Access
- Request access for custom pricing
Specifications
- Modalities
- Audio → Audio
- Access
- API (WebSocket)
- Released
- 2026-04-30 (Research announcement)
- Languages
- English (at launch)
- Architecture
- Asynchronous World Model
- Communication Mode
- Full-Duplex
- Transport Protocol
- WebSocket
- Input Modality
- Audio
- Output Modality
- Audio
- Status
- Beta
FAQ
What is Hydra?
Hydra is a native speech-to-speech model that operates as an asynchronous world model. It is designed for real-time, full-duplex voice experiences where the agent can listen, think, and speak concurrently.
When should I use Hydra?
You should use Hydra when you need true full-duplex conversations, natural barge-in capabilities, and low-latency interactions that preserve paralinguistic information like emotion and prosody.
How does Hydra handle communication?
Hydra uses a single WebSocket for audio-in and audio-out communication. This eliminates the need to orchestrate separate STT, LLM, and TTS models in a sequential pipeline.
Is Hydra available for public use?
Hydra is currently in beta. Interested developers can request access through the Smallest AI pricing page.
Added by AntSpace from public sources (smallest.ai).