Kokoro-82M VRAM & GPU Requirement Calculator
Interactive hardware sizing for Kokoro-82M (0.082B by hexgrad). Estimated 8K context footprint: 0.5 GB (INT4) • 0.5 GB (FP8) • 0.6 GB (FP16).
Non-autoregressive encoder/DiT (0 KV cache)
GPUs
Memory Requirements 0.59 GB
Requires 1 GPUs (based on memory capacity)
0.16 GB
All model weights
0 GB
Conversation history cache
0.37 GB
Concurrent processing cache
0.06 GB
Temporary computation cache
Performance Metrics
Visualize 5224 tok/s →Max Streams (This GPU)
236 streams
Duplex Architecture
Stream / Turn-Based
Estimated Latency
29 ms
Max QPS
34.8 queries/s
Estimated using memory bandwidth and tensor compute throughput across 1x NVIDIA RTX 4090 (24GB).
Quick Deploy CLI Command
docker run --gpus all -p 8880:8880 ghcr.io/remsky/kokoro-fastapi-gpu:latest
Flags reflect your selected precision, context length, and tensor parallel GPU count.Open Kokoro-82M spec page →
Kokoro-82M Architecture & Memory Profile
Total Parameters
0.082B
Architecture
Dense Transformer
Developer / Provider
hexgrad
Recommended Precision
FP16
Kokoro-82M VRAM Matrix by Precision & Concurrent Streams
Precomputed audio-stream VRAM requirements including unquantized acoustic/codec submodules and runtime workspace.
| Precision | 1 Stream | 4 Streams | 16 Streams |
|---|---|---|---|
| FP16 | 0.6 GB | 0.9 GB | 2.1 GB |
| FP8 | 0.5 GB | 0.8 GB | 2.0 GB |
| INT8 | 0.5 GB | 0.8 GB | 2.0 GB |
| INT4 | 0.5 GB | 0.8 GB | 2.0 GB |
Recommended GPU Hardware for Kokoro-82M (FP16)
Minimum GPU count and maximum concurrent real-time voice streams per node.
| GPU Hardware | Min GPUs | Total VRAM Pool | Max Streams |
|---|---|---|---|
| NVIDIA RTX 4090 (24GB) | 1× | 24 GB | 236 streams |
| NVIDIA RTX 5090 (32GB) | 1× | 32 GB | 317 streams |
| Apple M4 Max (128GB) (128GB) | 1× | 128 GB | 512 streams |
| NVIDIA A100 (80GB) (80GB) | 1× | 80 GB | 512 streams |
| NVIDIA H100 (80GB) | 1× | 80 GB | 512 streams |
| NVIDIA H200 (141GB) | 1× | 141 GB | 512 streams |
| Huawei Ascend 910B (64GB) | 1× | 64 GB | 512 streams |
Kokoro-82M Deployment FAQ
How much VRAM is needed to run Kokoro-82M?
For single-stream inference, Kokoro-82M uses about 0.6 GB in FP16, 0.5 GB in FP8, and 0.5 GB in INT4 including resident acoustic codecs and stream buffers.
Can I run Kokoro-82M on RTX 4090 or H100 GPUs?
Running Kokoro-82M in FP16 precision requires at least 1x RTX 4090 (24GB) or 1x NVIDIA H100 (80GB) GPUs.
What precision should I choose for Kokoro-82M in production?
For Kokoro-82M (0.082B by hexgrad), FP16 provides the best balance between inference speed, VRAM footprint, and output quality.