NVIDIA Nemotron VoiceChat: first open full-duplex model to call tools
One 11B model replaces the ASR-LLM-TTS stack, barges in at 450ms, runs on one GPU — but it's research-only. Plus: npm's supply chain under active siege.
Copy markdown
First open voice model that calls tools mid-sentence
NVIDIA's Nemotron VoiceChat 11B is the first open full-duplex speech-to-speech model that can fire tool calls while it's still talking. That means a self-hosted voice agent that looks things up or takes actions without breaking the conversation — and the weights are on Hugging Face now.
One 11B model, not an ASR-LLM-TTS pipeline
A single hybrid Mamba/Transformer (on a Nemotron Nano v2 9B backbone) does streaming speech-in and speech-out in one pass, with ~450ms turn-taking and instant barge-in when you interrupt. It serves on one A100/H100-class GPU via vLLM.
The catch: research-only, English-only
It ships under the OpenMDW v1.1 license for research use and handles English only. Prototype and benchmark it freely this weekend, but it's not cleared to drop into a production product yet.
The receipts: #2 on VoiceBench, 82.5% tool accuracy
It ranks #2 among open full-duplex models on VoiceBench and hits 82.5% tool-selection accuracy on Full-Duplex-Bench v3, with a 0.82 turn-taking score at 448ms latency — trained on ~550K hours of speech.
Elsewhere: npm's supply chain is under active attack
CrowdStrike's 2026 Threat Hunting Report pins 87% of software-registry threats this year on npm. North Korea's STARDUST CHOLLIMA slipped malware into 131 Mastra AI-framework packages, one actor poisoned 300+ dependencies in a day, and PoCs are weaponized within 48 hours. Audit your lockfiles.