---
license: other
license_name: nvidia-open-model-license
license_link: https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/
language:
- en
base_model:
- nvidia/personaplex-7b-v1
- kyutai/moshiko-pytorch-bf16
library_name: moshi
pipeline_tag: audio-to-audio
tags:
- personaplex
- moshi
- speech-to-speech
- full-duplex
- raw-weights
---

# PersonaPlex 7B v1 raw-weight mirror

This is a public, ungated, byte-identical mirror of the model artifacts from
[`nvidia/personaplex-7b-v1`](https://huggingface.co/nvidia/personaplex-7b-v1)
at immutable upstream revision
`fdaf4090a61cb315c138a1faee287ffd6c716309`.

This mirror exists to make reproducible deployment, quantization, adapter
training, and research automation possible without silently substituting the
historical `student_best.pt` derivative. It is independently maintained and is
not an NVIDIA repository.

## Integrity and provenance

| Artifact | Bytes | SHA-256 |
|---|---:|---|
| `model.safetensors` | 16,742,874,000 | `db1290db583cdaa6cb4de444ed279e0b586ca2a372b41434b07a7461c8c0e2f4` |
| `tokenizer-e351c8d8-checkpoint125.safetensors` | 384,644,900 | `09b782f0629851a271227fb9d36db65c041790365f11bbe5d3d59369cf863f50` |
| `tokenizer_spm_32k_3.model` | 552,778 | `78d4336533ddc26f9acf7250d7fb83492152196c6ea4212c841df76933f18d2d` |
| `voices.tgz` | 6,095,521 | `8564e9ca7a06ca723b07c3a77c623f0faa5937d04b2647b3a727b06c5ca0b7bb` |

`PROVENANCE.json` records the upstream revision and transformation status.
`UPSTREAM_SHA256SUMS` covers every mirrored upstream file. The original NVIDIA
model card is preserved as `UPSTREAM_README.md`.

## Use

```bash
pip install moshi
python -m moshi.server --hf-repo cudabenchmarktest/personaplex-7b-v1-raw-mirror
```

For the maintained PersonaPlex runtime and source instructions, use
[`NVIDIA/personaplex`](https://github.com/NVIDIA/personaplex).

## What this repository is not

- It is not the historical `cudabenchmarktest/personaplex-7b-nf4-distilled`
  checkpoint. That file is a BF16-derived training artifact with a different
  byte size and SHA-256.
- It is not an NF4 conversion.
- It contains no semantic-control adapter and makes no semantic-control quality
  claim.
- It has not modified, merged, pruned, quantized, or fine-tuned the raw weights.

## License

Use and redistribution are governed by the
[NVIDIA Open Model License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/).
The required NVIDIA attribution is in `NOTICE`, and copies of the governing
license materials are under `LICENSES/`. Additional base-model attribution is
in `THIRD_PARTY_NOTICES.md`.