Fun-CosyVoice3-0.5B VRAM & GPU Requirement Calculator

Interactive hardware sizing for Fun-CosyVoice3-0.5B (0.5B by FunAudioLLM). Estimated 8K context footprint: 2.5 GB (INT4) • 2.7 GB (FP8) • 3.2 GB (FP16).

Fixed 30s audio window / codec frame cap
GPUs

Memory Requirements 3.22 GB

Requires 1 GPUs (based on memory capacity)

2.1 GB

All model weights

0.02 GB

Conversation history cache

0.95 GB

Concurrent processing cache

0.14 GB

Temporary computation cache

Performance Metrics
Visualize 857 tok/s →
Max Streams (This GPU)
48 streams
Duplex Architecture
Stream / Turn-Based
Estimated Latency
175 ms
Max QPS
5.7 queries/s

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

Quick Deploy CLI Command
python runtime/python/fastapi/server.py --model_dir pretrained_models/Fun-CosyVoice3-0.5B
Flags reflect your selected precision, context length, and tensor parallel GPU count.Open Fun-CosyVoice3-0.5B spec page →

Fun-CosyVoice3-0.5B Architecture & Memory Profile

Total Parameters
0.5B
Architecture
Dense Transformer
Developer / Provider
FunAudioLLM
Recommended Precision
FP16

Fun-CosyVoice3-0.5B VRAM Matrix by Precision & Concurrent Streams

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

Precision1 Stream4 Streams16 Streams
FP163.2 GB4.5 GB9.8 GB
FP82.7 GB4.0 GB9.1 GB
INT82.7 GB4.0 GB9.1 GB
INT42.5 GB3.7 GB8.8 GB

Recommended GPU Hardware for Fun-CosyVoice3-0.5B (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 GB48 streams
NVIDIA RTX 5090 (32GB)1×32 GB66 streams
Apple M4 Max (128GB) (128GB)1×128 GB286 streams
NVIDIA A100 (80GB) (80GB)1×80 GB176 streams
NVIDIA H100 (80GB)1×80 GB176 streams
NVIDIA H200 (141GB)1×141 GB315 streams
Huawei Ascend 910B (64GB)1×64 GB139 streams

Fun-CosyVoice3-0.5B Deployment FAQ

How much VRAM is needed to run Fun-CosyVoice3-0.5B?
For single-stream inference, Fun-CosyVoice3-0.5B uses about 3.2 GB in FP16, 2.7 GB in FP8, and 2.5 GB in INT4 including resident acoustic codecs and stream buffers.
Can I run Fun-CosyVoice3-0.5B on RTX 4090 or H100 GPUs?
Running Fun-CosyVoice3-0.5B in FP16 precision requires at least 1x RTX 4090 (24GB) or 1x NVIDIA H100 (80GB) GPUs.
What precision should I choose for Fun-CosyVoice3-0.5B in production?
For Fun-CosyVoice3-0.5B (0.5B by FunAudioLLM), FP16 provides the best balance between inference speed, VRAM footprint, and output quality.