Verify the model and licensing
NVIDIA distributes code through its public PersonaPlex repository and publishes model access requirements. Accept the stated Hugging Face license before using model weights.
NVIDIA PersonaPlex: setup, deployment and GPU planning.
What it takes to turn a full-duplex speech model into a usable voice agent—without confusing a demo with a production calling service.
TECHNICAL REFERENCE · REVIEWED OCTOBER 10, 2026 · CHECK REPOSITORY RELEASES FOR CHANGES
01 / MODELWHAT PERSONAPLEX DOESOFFICIAL SOURCE
PersonaPlex is NVIDIA's research implementation of a speech-to-speech conversational model built on the Moshi architecture. It supports full-duplex exchanges—listening while speaking—with text-based role prompts and audio-based voice conditioning. Unlike a complete call-center platform, the released model is a building block. Developers must still handle communication services, business integration, identity, security and operations.
Start with the official NVIDIA/personaplex repository ↗. This page intentionally avoids claiming a fixed GPU-concurrency number because batch size, dtype, audio processing, memory pressure and latency targets differ by deployment.
02 / IMPLEMENTATIONSIX DEPLOYMENT CHECKPOINTSENGINEERING WORKFLOW
Install the audio dependencies
The official README lists libopus-dev on Ubuntu/Debian and installation of the moshi/ Python package. Align PyTorch and CUDA to your actual GPU; consult the current repository instructions.
Test local browser conversation
NVIDIA documents starting the server using python -m moshi.server with an SSL directory. Check real-time duplex interaction, interruptions, and audio round-trip quality before adding phone calls.
Profile memory and concurrency
A 7B-parameter model has material GPU-memory requirements. CPU offload is documented for low-memory environments, but it can add latency. Measure peak VRAM, latency and concurrent sessions on your target hardware rather than treating an untested GPU as a guaranteed fit.
Add telephony and business tools carefully
Connecting a phone provider, calendar, CRM or booking endpoint is application engineering: it is not automatically supplied by the research model. Keep authenticated actions behind a tool gateway and confirmation rules.
Run production readiness tests
Measure p50/p95 response delay, cross-talk, interruptions, reconnect behavior, telephony errors and rejected actions. Validate consent, privacy, and local calling requirements for the countries served.
03 / COMMANDSSTART FROM NVIDIA'S READMEEXAMPLE · UBUNTU
The commands below summarize the published install path. They assume a suitable Python/GPU environment, a clone of the NVIDIA repository and an approved model license. Commands and dependency versions can change.
sudo apt update && sudo apt install -y libopus-dev
# From the cloned NVIDIA/personaplex project:
pip install moshi/
# Authenticate to Hugging Face and accept the model license
SSL_DIR=$(mktemp -d)
python -m moshi.server --ssl "$SSL_DIR"
# If GPU memory is insufficient, NVIDIA documents --cpu-offload
# Install accelerate before trying offload.Do not expose a development server or short-lived SSL certificate directly to public users. Use appropriate TLS termination, access control, network isolation and monitoring for production deployments.
04 / CAPACITYGPU AND INFRASTRUCTURE QUESTIONSNO INVENTED BENCHMARKS
Can PersonaPlex run on a 16 GB GPU?
Do not assume that a 7B model will always fit at production latency on a 16 GB card. Parameter storage, runtime buffers, precision, KV/state caching, audio stages and framework overhead all matter. CPU offloading may allow experiments with less VRAM, at a possible performance cost.
Can one GPU serve 10, 50 or 100 concurrent callers?
The public repository does not establish a universal production concurrency guarantee. Measure sustained real-time-factor, GPU utilization, memory headroom, buffering, call-setup overhead and peak p95 response delay. Scale by verified session capacity and failover rather than by parameter count alone.
Will it automatically support Hindi, Twilio and customer appointments?
The released model's stated capabilities and prompts must be checked against language and use-case requirements. Phone providers and appointment APIs require separate integration work. Do not claim multilingual quality or reliable transaction completion until validated on real target-language evaluations.
Is CPU offload free GPU hosting?
No. It reduces memory pressure and trades compute/memory transfer for slower performance. A hosted service still needs servers, connectivity, logging and operations.
05 / READ MOREDETAILED GUIDESRECOUPREV RESEARCH
IN-DEPTH EXPLAINER
NVIDIA PersonaPlex Guide: Build and Scale AI Voice Agents
Install NVIDIA PersonaPlex-7B, connect live phone calls with Twilio, design secure booking workflows, plan GPU capacity and understand the limits of real-time voice AI.
READ THE GUIDE →IN-DEPTH EXPLAINER
AI Agents vs Chatbots: What Changes When Software Can Take Actions
An AI assistant answering a question and an AI agent changing a customer record may look similar on screen, but their risks and costs are fundamentally different.
READ THE GUIDE →IN-DEPTH EXPLAINER
AI Model Benchmarks Explained: Arena, SWE-bench and Intelligence Index
Learn how to verify 2026 AI model leaderboards using Arena human votes, Artificial Analysis scores and SWE-bench code evaluations without misreading rankings.
READ THE GUIDE →Before connecting real booking, messaging or business tools, review our AI-agent permissions checklist.
Primary reference: NVIDIA PersonaPlex source repository. This page offers an educational architecture overview rather than independently benchmarked GPU performance or a guarantee of commercial readiness. Editorial policy ↗.