Independent MLX/VLM MXFP8 conversion of meta-models/Muse-Glimmer-30B, pinned to source revision f84ecc3a0ea984a4c04542a84269e3d065350a6e. This is a community conversion and is not an official Meta release.

Format

Quantization: MXFP8, 8-bit, group size 32
Quantized scope: 416 language decoder projections
Preserved in BF16: normalized token embedding, untied LM head, vision tower, vision adapter, and vision projection
Indexed size: 32.76 GiB
Architecture: dense 52-layer Muse Glimmer language model with its 50-layer perception encoder
Tokenizer, chat template, processor metadata, license, and usage policy are preserved from the pinned source

The embedding and LM head intentionally remain BF16. Quantizing those terminal matrices caused deterministic EOS or special-token collapse in local tests. MXFP4 and MXFP8 were each produced directly from the same BF16 parent; neither was requantized from the other.

Runtime adapter

The release predates upstream muse_glimmer support in MLX-VLM 0.6.3. The exact adapter used for conversion and validation is bundled under runtime_adapter/muse_glimmer. Install it into an isolated environment before loading the model:

bash

python -m venv .venv
. .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install mlx==0.32.0 mlx-lm==0.31.3 mlx-vlm==0.6.3   torch==2.13.0 torchvision==0.28.0
python -m pip install   "transformers @ git+https://github.com/huggingface/transformers.git@a61d9a57c1ca1018fd84acabbf2104fdf468e143"

hf download Shiftedx/Muse-Glimmer-30B-MXFP8-Vision-MLX --local-dir ./Muse-Glimmer-30B-MXFP8
python - <<'PY'
from pathlib import Path
import shutil
import mlx_vlm

model_dir = Path("Muse-Glimmer-30B-MXFP8")
package_dir = Path(mlx_vlm.__file__).resolve().parent / "models" / "muse_glimmer"
if package_dir.exists():
    raise SystemExit(f"Refusing to replace existing adapter: {package_dir}")
shutil.copytree(model_dir / "runtime_adapter" / "muse_glimmer", package_dir)
print("Installed the bundled Muse Glimmer adapter.")
PY

Then generate with MLX-VLM:

python -m mlx_vlm.generate   --model ./Muse-Glimmer-30B-MXFP8   --image image.jpg   --prompt "Describe this image."   --max-tokens 256

Qualification status

Publisher checkpoint parameter tree: exact 1,436-tensor match
Adapter reference and round-trip fixtures: pass
Structural inspection and all shard headers: pass
Strict, non-lazy MLX-VLM load: pass
Basic deterministic text generation: pass
Structured synthetic-image color recognition: pass

This artifact has not completed broad BF16 parent-parity, agentic, tool-use, long-context, or real-image qualification. Treat it as an experimental build and evaluate it for your workload before deployment. Generated text may expose the model's reasoning/template structure depending on the calling scaffold.

Scope and limitations

This repository does not contain the publisher's DFlash companion and is not an MTPLX package. Muse Glimmer's DFlash block-diffusion assistant is not an MTP sidecar. Use a runtime with an explicit Muse Glimmer/DFlash backend if speculative decoding is required.

Review the source model card and the included USAGE_POLICY.md before use. Quantization does not change the source model's license, usage restrictions, or safety characteristics.