From idea toprototypein minutes
Real-time speech models with 50 ms latency, an open-source framework for prototyping and orchestration partners for production. Pick your stack and go.
47ms
Time to first audio (P50)
5.3%
Word error rate
#1
Leading silence on Coval benchmarks
Open-source framework
Prototype
with Gradbot
A small open-source framework for testing Gradium models in a live agent before you write any plumbing.
What is Gradbot
Build a voice agent in a single prompt. Gradbot keeps STT, LLM and TTS in sync so you only focus on the logic. Works with any OpenAI-compatible LLM.
- Browse demos
Demos
Runnable agents to start from instead of a blank file. Clone one, swap the prompt, and ship.
- Add the skill
Coding-agent skill
A skill that teaches your coding agent Gradbot's patterns, so it scaffolds and extends voice agents correctly on the first try.
- Read the post
Why we built it
The problems we hit prototyping voice agents, and the design decisions behind Gradbot's core.
Works with
Everything for builders
Run Gradium in production on the stack you already use.
Docs & ressources
Everything else
you need
When you’re ready for production scale, Gradium models run natively in the orchestration frameworks you already use.
Documentation
Quickstarts, the Python SDK, streaming TTS, STT with semantic VAD, and the full REST and WebSocket reference.
Read the docsMigration guide
Coming from ElevenLabs? The migration guide is the most-read page in our docs for a reason.
Read the migration guideExamples repo
Voice agent demos beyond Gradbot: integrations, on-device experiments, and reference apps.
Explore GitHubTechnical deep dives
How we measure time to first audio, how semantic VAD works, and how we trade quality against latency.
Read the blog
Community
Don’t
build alone
Gradbot is open source and shaped by the people using it.
