NVIDIA's PersonaPlex is a full-duplex speech model that processes audio continuously and naturally, running entirely on consumer hardware without cloud dependencies. Released in January 2026, this 7-billion-parameter model from NVIDIA represents a fundamental shift in how conversational AI works. Unlike traditional voice assistants that transcribe speech to text, send it to a language model, and then convert the response back to audio, PersonaPlex thinks in voice from start to finish.

What Makes Full-Duplex Speech Different From Regular Voice Assistants?

Most AI assistants today offer what feels like voice interaction, but it's actually a multi-step process layered on top of text-based models. You speak, the system transcribes your words, a language model processes the text, and then text-to-speech software reads the response aloud. This approach introduces latency and loses the natural flow of human conversation.

PersonaPlex works fundamentally differently. It continuously processes two audio streams simultaneously: your voice and its own response, while also predicting what it will say next. This mirrors how humans actually converse. When someone speaks to you, you don't write down their words, analyze them, and then compose a written response. You think in voice, and PersonaPlex does the same.

The model is built on Moshi, an open-source full-duplex speech model from Kyutai. Underneath, it uses a neural audio codec called Mimi that compresses speech into tokens at 12.5 frames per second. A large temporal transformer then models the conversation flow, while a smaller depth transformer handles the audio codes within each frame.

How Does PersonaPlex Run on Consumer Hardware?

What makes PersonaPlex particularly noteworthy is its accessibility. NVIDIA fine-tuned the Moshi foundation to allow controllable roles and voices. You provide a text prompt describing who the model should be, plus a short voice sample, and the model adapts accordingly. Demo personas include a teacher, a customer-service agent, and Alex, an astronaut character.

Setup is straightforward for those comfortable with command-line tools. You need the repository, PyTorch, the Opus audio library, access to the model weights on Hugging Face, and a browser with microphone permission. The model downloads at approximately 9 gigabytes in size. Once downloaded, the entire conversation stays on your machine with no API costs or per-minute billing.

On an NVIDIA RTX 3090 graphics card, the standard BF16 implementation uses around 19 gigabytes of video RAM and runs smoothly in real time. That's substantial but hardly supercomputer territory. A used RTX 3090 can run an end-to-end conversational speech model that would have seemed like a research demo from the distant future just a few years ago.

Can You Run PersonaPlex on Smaller GPUs?

Community projects have already extended PersonaPlex's reach to more modest hardware. Projects like the PersonaPlex ComfyUI integration and moshi.cpp have added 8-bit and 4-bit quantization support, enabling the model to run on graphics cards with 8 to 16 gigabytes of memory. Some reports indicate strong performance even on an 8-gigabyte RTX 2070 laptop GPU, though performance depends on the quantization method, backend software, and how much computation spills onto the CPU.

What Are the Current Limitations of PersonaPlex?

PersonaPlex is impressive in real-time conversation, but it has meaningful constraints. The model maintains a rolling memory of 3,000 frames. At 12.5 frames per second, this translates to approximately four minutes of conversation history. Once the memory fills, new frames overwrite the oldest ones, causing the model to forget earlier parts of the conversation.

This limitation became apparent during extended testing. A persona named Alex, an astronaut, gradually lost her identity after four minutes of conversation about unrelated topics like money and cars. When asked her name, Alex suddenly declared, "My name is Kim. I'm actually a very good friend of yours." The model had simply forgotten its original persona because that information had fallen out of the context window.

The deterioration accelerated from there. Responses became repetitive, with phrases like "That's very good" morphing into "Very, very good for you," and eventually the model became trapped in a loop, repeatedly asking "What's going to happen?" The current implementation lacks an effective repetition penalty to pull the model out of such loops.

How to Extend PersonaPlex's Conversational Memory

• Increase the Attention Cache: The 3,000-frame attention cache costs roughly 1.46 gigabytes of video RAM and grows linearly, so an eight-minute cache would be plausible on a 24-gigabyte card, though this only postpones the memory problem rather than solving it.
• Optimize for Larger Context Windows: Expanding the attention cache is technically possible, but inference speed may suffer despite having more powerful hardware, and the model was designed to operate at the default 3,000-frame setting.
• Use Community Quantization Tools: Projects like moshi.cpp and ComfyUI integrations allow you to run PersonaPlex on smaller GPUs, which may help manage memory constraints for longer conversations on more modest hardware.

Increasing the context window addresses the symptom but not the underlying architecture. The model's designers intended it to operate with a four-minute memory window, and pushing beyond that may introduce performance trade-offs that outweigh the benefits.

What Does PersonaPlex Mean for the Future of Local AI?

PersonaPlex demonstrates that sophisticated conversational AI no longer requires cloud infrastructure or API subscriptions. The model runs entirely locally, processes audio natively without text conversion, and adapts to different personas and voices. For users concerned about privacy or those without reliable internet, this represents a significant milestone.

The four-minute memory limitation is a real constraint for long conversations, but it's also a solvable engineering problem. Future iterations could implement longer context windows, more effective repetition penalties, or hybrid approaches that maintain longer-term memory without proportionally increasing computational costs. The fact that community developers have already optimized PersonaPlex for smaller GPUs suggests that the open-source ecosystem will continue pushing these boundaries.

PersonaPlex also highlights the broader shift toward full-duplex speech models as the natural interface for conversational AI. OpenAI's ChatGPT live mode operates on similar principles, processing audio streams continuously rather than converting to text as an intermediate step. As these models mature and become more efficient, they may reshape expectations for what "natural" AI conversation actually means.