BREAKING
LiveKit Tunes Gemma 4 31B for Voice
0ms
Time to first token
0ms
Time to first audio
Tool Use vs Task Completion
tau2bench76.9
Hotel tasks done88
Drop-In Voice Pipeline
1Deepgram STT
2Gemma 4 31B
3Cartesia TTS
Hosted or Self-Host
LiveKit InferenceHosted
$1.20 per million tokens
Python and TypeScript SDKs
Self-HostOpen-weight
Apache 2.0 license
On Hugging Face and AWS Bedrock
Speed Meets Tool Use
AI NEWS BLITZ
LiveKit has launched a latency-optimized Gemma 4 31B for real-time voice AI agents.