Keye-VL-2.0-30B-A3B — Abliterated — MLX (mixed ~3.87 bpw)

Mixed-precision MLX build (≈3.87 bpw) of the abliterated Kwai-Keye/Keye-VL-2.0-30B-A3B, for Apple Silicon. Quantized from lemuralabs/Keye-VL-2.0-30B-A3B-uncensored with mlx-vlm's mixed_3_4 recipe (3/4-bit mix, group size 64). ~14 GB on disk.

This is the "optiq ~3.7 bpw" target build. The closest standard mlx-vlm mixed recipe to 3.7 bpw is mixed_3_4, which measured 3.871 bpw (mixed_3_6 was 4.22). If you need a smaller footprint, see the mxfp4 build (≈4.43 bpw but a different format) or request a custom predicate.

Yes — Like the mxfp4 build, this runs Keye coherently on Mac via dense attention (the stock CUDA SALightningIndexer is unstable on MPS; this port drops it).

Requirements & usage

Same as the mxfp4 build: this needs the bundled mlx-vlm support module in mlx_vlm_keye_support/keyevl2/ (copy into mlx_vlm/models/keyevl2/, and add "keye_vl2": MessageFormat.LIST_WITH_IMAGE_FIRST, to mlx_vlm/prompt_utils.py). Then:

python -m mlx_vlm generate --model lemuralabs/Keye-VL-2.0-30B-A3B-uncensored-mlx-optiq-3.7bpw \
 --prompt "Describe this image." --image path/to/img.jpg --trust-remote-code

Notes

• Quant: mlx-vlm mixed_3_4, group size 64, 3.871 bpw (whole model incl. vision).
• Vision: SigLIP tower + mlp_AR projector included; text gen verified coherent, image-understanding forward is a first-cut port.
• Abliterated; text backbone reuses mlx-vlm qwen3_vl_moe, sparse sa_indexer dropped.

Abliteration removes safety alignment; you are responsible for use.

Other variants of this model (public on Lemura Labs)

• Keye-VL-2.0-30B-A3B-uncensored — abliterated bf16 — full precision (CUDA / vLLM / SGLang)
• Keye-VL-2.0-30B-A3B-uncensored-mlx-mxfp4 — MLX MXFP4 (~4.43 bpw) — Apple Silicon