Qwen3.8-27B-4bit

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. The 333 vision_tower.* tensors and multimodal processor assets come from mlx-community/Qwen3.8-27B-4bit@3e6447f082e89cc7f0bc6e5441afd38dfce760ff, converted from the official Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 source. Tensor dtype, shape, and raw bytes were checked before publication.

The immutable challenge revision 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

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:

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.