Modelpublic

bsisduck/Muse-Glimmer-30B-MLX-MXFP8

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes53downloads

Model Card

Muse Glimmer 30B MLX MXFP8

Staging status: tensor integrity is verified, but inference did not complete on the tested 32 GB host because Metal ran out of memory. This artifact is prepared for validation on a higher-memory Apple Silicon host and is not release-ready yet.

An MLX-VLM MXFP8 conversion of `meta-models/Muse-Glimmer-30B`, developed by Meta Superintelligence Lab. The conversion is pinned to source revision 97c77dff50b2797bcc558fa2d909761dbc575c59.

Model details

Property	Value
Architecture	Dense causal transformer with a dedicated perception encoder
Parameters	Approximately 29.8B total, including an approximately 1.8B vision encoder
Modalities	Text and image input; text output
Context length	131,072+ architecture limit documented upstream; not validated here
Quantization	RTN `mxfp8`, 8-bit, group size 32
Quantization scope	Eligible language and bridge linear layers; `vision_tower` remains unquantized
Artifact size	Approximately 32.53 GB
Release status	Staging; higher-memory inference validation required

Requirements for the next test

—Apple Silicon with 64 GB or more unified memory is the next recommended validation target. This is a test recommendation, not a verified minimum.
—Approximately 34 GB of disk space for the model, plus runtime and swap space.
—macOS, Python 3.13, and the pinned MLX-VLM runtime below.

bash

pip install "mlx-vlm @ git+https://github.com/Blaizzy/mlx-vlm.git@132fc5dea46a085f39400cb101d1ad7c1dabdd49"

Test commands

Text

bash

mlx_vlm.generate \
  --model bsisduck/Muse-Glimmer-30B-MLX-MXFP8 \
  --prompt "Return only the result of 84 * 3 / 2." \
  --max-tokens 48 --temperature 0

Image

bash

mlx_vlm.generate \
  --model bsisduck/Muse-Glimmer-30B-MLX-MXFP8 \
  --image ./image.jpg \
  --prompt "Describe this image and identify the important details." \
  --max-tokens 128 --temperature 0

Validation status

Check	Result
Conversion	Passed; bounded 2 GiB serialization, 30.06 seconds
Serialized tensor comparison	Passed; 1857/1857 tensors exactly match the pinned MXFP8 reference
Local-to-remote transfer	Passed; 24 files match SHA-256 manifest digest `65ec5c8d4e12879e397f9f662cb6410360b5a4e4ade24e01b9afd3badeb9e25b`
Text smoke on 32 GB	Not completed; Metal out of memory
Vision smoke on 32 GB	Not started because the text gate did not pass

The failed smoke attempt reported a 30997 MB model requirement against a 25559 MB recommended runtime limit. It does not indicate corrupted weights: the independent tensor audit matched every serialized tensor.

Limitations and safety

—No inference-quality, throughput, long-context, or DFlash claim is made for this staging artifact.
—Quantization can change model outputs relative to BF16 even when serialized tensors exactly match another MXFP8 conversion.
—The model can produce inaccurate, biased, unsafe, or objectionable output.
—Read the upstream model card and `USAGE_POLICY.md` before deployment.

Staging evidence

—`reports/conversion.json`
—`reports/tensor_audit.json`
—`reports/remote_validation.json`
—`reports/runtime.json`