You can deploy this model and start earning money today!

lemuralabs/
Keye-VL-2.0-30B-A3B-uncensored-mlx-mxfp4

Name: Keye-VL-2.0-30B-A3B-uncensored-mlx-mxfp4
Availability: InStock

Kwai-Keye/Keye-VL-2.0-30B-A3B

quantized

apache-2.0

4-bit

Views

No views yet

Lemura Labs

Keye-VL-2.0-30B-A3B — Abliterated — MLX (MXFP4)

MXFP4 (≈4.43 bpw) MLX build 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. ~16 GB on disk; runs in ~17 GB; ≈107 tok/s on an M4 Max.

Yes — This MLX build runs Keye coherently on a Mac. The original model uses a CUDA-only sparse-attention indexer (SALightningIndexer) that is unstable on MPS — so the stock model generates garbage via Transformers on Apple Silicon. This port runs the mathematically-equivalent dense attention, which is coherent and fast on MLX.

Requirements — custom mlx-vlm model class

Keye is not yet in mainline mlx-vlm. This repo bundles the support module under mlx_vlm_keye_support/keyevl2/. Install it:

bash

pip install mlx-vlm
# copy the bundled module into your mlx-vlm install:
python - <<'PY'
import mlx_vlm, os, shutil
dst = os.path.join(os.path.dirname(mlx_vlm.__file__), "models", "keyevl2")
shutil.copytree("mlx_vlm_keye_support/keyevl2", dst, dirs_exist_ok=True)
print("installed keyevl2 ->", dst)
PY

Also register the prompt format (one line in mlx_vlm/prompt_utils.py): add "keye_vl2": MessageFormat.LIST_WITH_IMAGE_FIRST, to the format map.

Usage

bash

python -m mlx_vlm generate --model lemuralabs/Keye-VL-2.0-30B-A3B-uncensored-mlx-mxfp4 \
 --prompt "Describe this image." --image path/to/img.jpg --trust-remote-code

Notes

Quant: MXFP4, group size 32, 4.432 bpw (whole model, incl. vision tower).
Vision: the SigLIP tower + mlp_AR projector are included (quantized). Text gen is verified coherent; image understanding is functional but the packed-vision forward in this port is a first cut — report issues.
Abliterated (refusals reduced; see the base abliterated card for method/limits).
The text backbone reuses mlx-vlm's qwen3_vl_moe; the sparse sa_indexer is 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-optiq-3.7bpw — MLX mixed ~3.87 bpw — Apple Silicon

lemuralabs/Keye-VL-2.0-30B-A3B-uncensored-mlx-mxfp4