People expect near-instant, consistent responses from voice agents — which puts unique pressure on latency and correct tool invocation. This model is a deliberate trade: much smaller active compute per request than frontier models, but fine-tuned to call tools accurately and meet strict voice-agent time budgets without relying on "thinking" tokens.
Key Capabilities
- Low-latency, phone-agent-focused tuning: a full-parameter fine-tune of Nemotron 3 Nano that prioritizes quick time-to-first-token and short voice-to-voice latency so the LLM fits a ~650ms LLM budget in optimized stacks.
- Tool and function-calling discipline: trained to invoke the right tools at the right turn (say/do consistency) even with thinking disabled, reducing mismatches where the model claims actions it didn't perform.
- Efficient MoE architecture: hybrid Mamba-Transformer mixture-of-experts with 30B total parameters and ~3.5B active parameters, enabling high concurrency and low cost on modern NVIDIA hardware.
- Production-ready serving recommendations: bfloat16 safetensors, long context support (262,144 tokens), and guidance for vLLM/SGLang/Modal deployments; recommended inference settings: temperature=0 and thinking disabled.
Who it's for + tradeoffs
Great fit if you need a self-hostable voice agent LLM that minimizes per-minute inference cost and P95 latency while behaving reliably in multi-turn, tool-enabled customer-service flows. It suits teams that can run Nemotron-compatible inference stacks (vLLM, SGLang, Modal) and who prefer controllable deployments. Look elsewhere if you require multilingual coverage beyond English, broader general-purpose reasoning with chain-of-thought enabled, or if you cannot operate the GPU infrastructure needed for a 30B-weight model; this model is optimized for English phone-agent workloads and expects the deployment conventions (temperature=0, thinking disabled) it was trained with.
