Text-to-Speech
Safetensors
MLX
gbx-lm
fish_drama3
multi-speaker
voice-cloning
apple-silicon
expert-paging
mixture-of-experts
Instructions to use GreenBitAI/Drama-3-8bit-paged with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use GreenBitAI/Drama-3-8bit-paged with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] hf download GreenBitAI/Drama-3-8bit-paged --local-dir Drama-3-8bit-paged
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
Download experts_index.json from GreenBitAI/Drama-3-8bit-paged: direct link, hf CLI and curl.
- Browser
- Download file 3.02 kB
-
https://huggingface.co/GreenBitAI/Drama-3-8bit-paged/resolve/main/experts_index.json
- Command line
-
hf download hf://GreenBitAI/Drama-3-8bit-paged/experts_index.json
-
curl -L -o experts_index.json https://huggingface.co/GreenBitAI/Drama-3-8bit-paged/resolve/main/experts_index.json
3.02 kB
| { | |
| "format": "gbx-paged-experts-v1", | |
| "source": "", | |
| "num_layers": 40, | |
| "num_experts": 256, | |
| "layers": [ | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9, | |
| 10, | |
| 11, | |
| 12, | |
| 13, | |
| 14, | |
| 15, | |
| 16, | |
| 17, | |
| 18, | |
| 19, | |
| 20, | |
| 21, | |
| 22, | |
| 23, | |
| 24, | |
| 25, | |
| 26, | |
| 27, | |
| 28, | |
| 29, | |
| 30, | |
| 31, | |
| 32, | |
| 33, | |
| 34, | |
| 35, | |
| 36, | |
| 37, | |
| 38, | |
| 39 | |
| ], | |
| "fields": [ | |
| { | |
| "proj": "gate_proj", | |
| "field": "weight", | |
| "dtype": "uint32", | |
| "shape": [ | |
| 512, | |
| 512 | |
| ], | |
| "offset": 0, | |
| "nbytes": 1048576 | |
| }, | |
| { | |
| "proj": "gate_proj", | |
| "field": "scales", | |
| "dtype": "bfloat16", | |
| "shape": [ | |
| 512, | |
| 32 | |
| ], | |
| "offset": 1048576, | |
| "nbytes": 32768 | |
| }, | |
| { | |
| "proj": "gate_proj", | |
| "field": "biases", | |
| "dtype": "bfloat16", | |
| "shape": [ | |
| 512, | |
| 32 | |
| ], | |
| "offset": 1081344, | |
| "nbytes": 32768 | |
| }, | |
| { | |
| "proj": "up_proj", | |
| "field": "weight", | |
| "dtype": "uint32", | |
| "shape": [ | |
| 512, | |
| 512 | |
| ], | |
| "offset": 1114112, | |
| "nbytes": 1048576 | |
| }, | |
| { | |
| "proj": "up_proj", | |
| "field": "scales", | |
| "dtype": "bfloat16", | |
| "shape": [ | |
| 512, | |
| 32 | |
| ], | |
| "offset": 2162688, | |
| "nbytes": 32768 | |
| }, | |
| { | |
| "proj": "up_proj", | |
| "field": "biases", | |
| "dtype": "bfloat16", | |
| "shape": [ | |
| 512, | |
| 32 | |
| ], | |
| "offset": 2195456, | |
| "nbytes": 32768 | |
| }, | |
| { | |
| "proj": "down_proj", | |
| "field": "weight", | |
| "dtype": "uint32", | |
| "shape": [ | |
| 2048, | |
| 128 | |
| ], | |
| "offset": 2228224, | |
| "nbytes": 1048576 | |
| }, | |
| { | |
| "proj": "down_proj", | |
| "field": "scales", | |
| "dtype": "bfloat16", | |
| "shape": [ | |
| 2048, | |
| 8 | |
| ], | |
| "offset": 3276800, | |
| "nbytes": 32768 | |
| }, | |
| { | |
| "proj": "down_proj", | |
| "field": "biases", | |
| "dtype": "bfloat16", | |
| "shape": [ | |
| 2048, | |
| 8 | |
| ], | |
| "offset": 3309568, | |
| "nbytes": 32768 | |
| } | |
| ], | |
| "payload_bytes": 3342336, | |
| "record_bytes": 3342336, | |
| "align": 4096, | |
| "region": "experts.bin", | |
| "created": "2026-10-28T04:19:47", | |
| "trunk": "model.safetensors", | |
| "self_contained": true, | |
| "source_prefix": "backbone.model", | |
| "source_id": null, | |
| "ple_store": null, | |
| "engram_stores": null, | |
| "verified": { | |
| "against": "the source checkpoint", | |
| "at": "2026-10-27T20:20:07+00:00", | |
| "slots_per_layer": 103, | |
| "gates": [ | |
| { | |
| "gate": "gate 3", | |
| "name": "bit-identical logits", | |
| "status": "PASS", | |
| "detail": "exact on 4 prompt(s) to 127 tokens; 2 longer differ by at most 3, same greedy token throughout" | |
| }, | |
| { | |
| "gate": "gate 32", | |
| "name": "layer-wise vs resident", | |
| "status": "PASS", | |
| "detail": "40 layers x 2 draws exact against the checkpoint, 48.04 GiB peak for this gate" | |
| } | |
| ] | |
| } | |
| } |