Muse Glimmer 30B MLX BF16
Staging status: tensor integrity is verified, but inference was not attempted on the available 32 GB host because the artifact alone exceeds its unified memory. This artifact is prepared for validation on a higher-memory Apple Silicon host and is not release-ready yet.
An MLX-VLM BF16 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 |
| Precision | BF16 (bfloat16); not quantized |
| Precision scope | All floating-point model weights use BF16 |
| Artifact size | Approximately 59.58 GB |
| Release status | Staging; higher-memory inference validation required |
Requirements for the next test
- Apple Silicon with 96 GB or more unified memory is the next recommended validation target. This is a test recommendation, not a verified minimum.
- Approximately 61 GB of disk space for the model, plus runtime and swap space.
- macOS, Python 3.13, and the pinned MLX-VLM runtime below.
pip install "mlx-vlm @ git+https://github.com/Blaizzy/mlx-vlm.git@132fc5dea46a085f39400cb101d1ad7c1dabdd49"
Test commands
Text
mlx_vlm.generate \
--model bsisduck/Muse-Glimmer-30B-MLX-BF16 \
--prompt "Return only the result of 84 * 3 / 2." \
--max-tokens 48 --temperature 0
Image
mlx_vlm.generate \
--model bsisduck/Muse-Glimmer-30B-MLX-BF16 \
--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, 52.17 seconds |
| Serialized tensor comparison | Passed; 1436/1436 tensors exactly match the pinned BF16 reference |
| Local-to-remote transfer | Not applicable; artifact was not copied to the undersized host |
| Text smoke on 32 GB | Not attempted; available host is too small |
| Vision smoke on 32 GB | Not attempted; available host is too small |
The approximately 59.58 GB BF16 weights exceed the host's 32 GB unified memory before runtime and KV-cache overhead.
Limitations and safety
- No inference-quality, throughput, long-context, or DFlash claim is made for this staging artifact.
- No task-quality comparison against the original runtime was run; exact serialized tensors establish conversion integrity, not benchmark quality.
- The model can produce inaccurate, biased, unsafe, or objectionable output.
- Read the upstream model card and
USAGE_POLICY.mdbefore deployment.