File size: 2,422 Bytes
eda45ab a210166 2f330b9 eda45ab 2f330b9 eda45ab a210166 2f330b9 a210166 301e9e2 efd4e8c 2f330b9 a210166 2f330b9 a210166 2f330b9 a210166 2f330b9 301e9e2 2f330b9 a210166 2f330b9 a210166 2f330b9 a210166 2f330b9 a210166 2f330b9 a210166 2f330b9 a210166 2f330b9 | 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 40 41 42 43 44 45 46 47 | ---
library_name: mlx
license: apache-2.0
pipeline_tag: image-text-to-text
base_model:
- Qwen/Qwen3.8-27B
tags:
- mlx
- mlx-vlm
- vision
---
# EigenLabs/Qwen3.8-27B-4bit
This revision is a complete MLX 4-bit Qwen3.8 vision-language model that accepts image, video, and text inputs and generates text.
The **1,847 `language_model.*` tensors are bit-for-bit identical** to [`EigenLabs/Qwen3.8-27B-4bit@eda45ab47f465d08d6558f0353a2346e2eb9d5b3`](https://huggingface.co/EigenLabs/Qwen3.8-27B-4bit/tree/eda45ab47f465d08d6558f0353a2346e2eb9d5b3). The 333 `vision_tower.*` tensors and multimodal processor assets come from [`mlx-community/Qwen3.8-27B-4bit@3e6447f082e89cc7f0bc6e5441afd38dfce760ff`](https://huggingface.co/mlx-community/Qwen3.8-27B-4bit/tree/3e6447f082e89cc7f0bc6e5441afd38dfce760ff), converted from the official [`Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`](https://huggingface.co/Qwen/Qwen3.8-27B/tree/1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0) source. Tensor dtype, shape, and raw bytes were checked before publication.
The immutable challenge revision [`eda45ab47f465d08d6558f0353a2346e2eb9d5b3`](https://huggingface.co/EigenLabs/Qwen3.8-27B-4bit/tree/eda45ab47f465d08d6558f0353a2346e2eb9d5b3) remains available and is text-only because it has no vision tower or processor assets. Multi-token prediction (MTP) is intentionally **not embedded** in this repository; the MTP artifact remains separate.
## mlx-vlm
```bash
pip install -U mlx-vlm
mlx_vlm.generate \
--model EigenLabs/Qwen3.8-27B-4bit \
--image /path/to/image.jpg \
--prompt "Describe this image." \
--max-tokens 256
```
Pin the immutable Hub commit shown in this model's history for reproducible deployment.
## Darkbloom
This is the Hugging Face source artifact consumed by Darkbloom's in-process MLX VLM runtime. Darkbloom providers deploy only catalog-approved, immutable model revisions and verify their manifests before loading. Updating this repository alone does not register the model in the Darkbloom catalog or activate provider traffic.
After a Darkbloom catalog release pins the immutable Hub commit and publishes its verified manifest, a provider can discover and download that release:
```bash
darkbloom models catalog
darkbloom models download <catalog-model-id>
darkbloom start --foreground
```
`<catalog-model-id>` is assigned by the Darkbloom catalog; it is not the Hugging Face repository name.
|