personaplex-24L-q4_k-webgpu

The pruned, QLoRA-recovered, Q4_K-quantized PersonaPlex-7B that sts-web's live demo actually loads.

nvidia/personaplex-7b-v1, by NVIDIA

Pruned, QLoRA-recovered, Q4_K quantization for WebGPU by Idle Intelligence.

NVIDIA Open Model License

NVIDIA PersonaPlex-7B-v1 with 8 middle temporal-transformer layers removed, quality recovered with a rank-32 LoRA trained on self-distilled teacher audio, then quantized to Q4_K for browser WebGPU inference. Quality assessed only by listening tests so far. Not affiliated with or endorsed by NVIDIA.

Type
model
Runs
WebGPU
Size
6.74B params (24 of 32 temporal layers), 3.5GB (Q4_K)
Demo (GitHub Pages)
https://idle-intelligence.github.io/sts-web/web/
Source
idle-intelligence/sts-web
HuggingFace
idle-intelligence/personaplex-24L-q4_k-webgpu
Status
experiment

Used on: sts

← Home