From the source
Baseten is launching the Baseten Voice Agent Stack and Baseten Realtime Voice, both in private preview, to let teams run, customize and scale production voice agents on its own infrastructure.
- The stack runs the orchestrator and every model over a tight in-cluster network on Baseten Dedicated Inference, scales each component independently, and lets teams pick or post-train any model in the pipeline.
- Baseten Realtime Voice is a reference agent built on the stack that runs a cascade of Baseten-optimized open-weight speech models with a custom Pipecat-based orchestrator, and it speaks the
OpenAI Realtime protocol so existing agents can move over with minimal changes.
- On Baseten's setup it scores 93% on Full-Duplex-Bench, which Baseten says is the highest of any open-source system on Artificial Analysis, and 73.5% on τ²-Bench Voice, more than double the cascaded baseline.
- Baseten says it costs less per minute ($0.012) than every hosted speech-to-speech API it evaluated at 32+ concurrent streams, and that the reference agent achieves sub-second median turn latency across 32 concurrent sessions.
- The stack supports optional delegation, pairing a co-located talker model with a stronger thinker model for reasoning and tool calls, and server-side tool and MCP execution on the orchestrator to cut tool-calling latency.





