PersonaPlex-7B / Moshi (7.2B) VRAM & GPU Requirement Calculator

Interactive hardware sizing for PersonaPlex-7B / Moshi (7.2B) (7.2B by NVIDIA / Kyutai). Estimated 8K context footprint: 6.1 GB (INT4) • 9.7 GB (FP8) • 16.9 GB (FP16).

3,000-step 12.5Hz ring buffer (persistent FP16 KV)
GPUs

Memory Requirements 16.89 GB

Requires 1 GPUs (based on memory capacity)

14.65 GB

All model weights

1.46 GB

Conversation history cache

0.67 GB

Concurrent processing cache

0.1 GB

Temporary computation cache

Performance Metrics
Visualize 60 tok/s →
Max Streams (This GPU)
4 streams
Duplex Architecture
Native Full-Duplex (12.5Hz)
Estimated Latency
2521 ms
Max QPS
0.4 queries/s

Estimated using memory bandwidth and tensor compute throughput across 1x NVIDIA RTX 4090 (24GB).

Quick Deploy CLI Command
python -m moshi.server --hf-repo nvidia/personaplex-7b-v1
Flags reflect your selected precision, context length, and tensor parallel GPU count.Open PersonaPlex-7B spec page →

PersonaPlex-7B / Moshi (7.2B) Architecture & Memory ProfileTrue Native Full-Duplex (12.5Hz Lock)

Total Parameters
7.2B
Architecture
Dense Transformer
Developer / Provider
NVIDIA / Kyutai
Recommended Precision
FP16

PersonaPlex-7B / Moshi (7.2B) VRAM Matrix by Precision & Concurrent Streams

Precomputed audio-stream VRAM requirements including unquantized acoustic/codec submodules and runtime workspace.

Precision1 Stream4 Streams16 Streams
FP1616.9 GB21.8 GB41.6 GB
FP89.7 GB14.6 GB34.4 GB
INT89.7 GB14.6 GB34.4 GB
INT46.1 GB11.0 GB30.8 GB

Recommended GPU Hardware for PersonaPlex-7B / Moshi (7.2B) (FP16)

Minimum GPU count and maximum concurrent real-time voice streams per node.

GPU HardwareMin GPUsTotal VRAM PoolMax Streams
NVIDIA RTX 4090 (24GB)1×24 GB4 streams
NVIDIA RTX 5090 (32GB)1×32 GB5 streams
Apple M4 Max (128GB) (128GB)1×128 GB3 streams
NVIDIA A100 (80GB) (80GB)1×80 GB4 streams
NVIDIA H100 (80GB)1×80 GB24 streams
NVIDIA H200 (141GB)1×141 GB24 streams
Huawei Ascend 910B (64GB)1×64 GB7 streams

PersonaPlex-7B / Moshi (7.2B) Deployment FAQ

How much VRAM is needed to run PersonaPlex-7B / Moshi (7.2B)?
For single-stream inference, PersonaPlex-7B / Moshi (7.2B) uses about 16.9 GB in FP16, 9.7 GB in FP8, and 6.1 GB in INT4 including resident acoustic codecs and stream buffers.
Can I run PersonaPlex-7B / Moshi (7.2B) on RTX 4090 or H100 GPUs?
Running PersonaPlex-7B / Moshi (7.2B) in FP16 precision requires at least 1x RTX 4090 (24GB) or 1x NVIDIA H100 (80GB) GPUs.
What precision should I choose for PersonaPlex-7B / Moshi (7.2B) in production?
For PersonaPlex-7B / Moshi (7.2B) (7.2B by NVIDIA / Kyutai), FP16 provides the best balance between inference speed, VRAM footprint, and output quality.