bsisduck/Muse-Glimmer-30B-MLX-MXFP4

Muse Glimmer 30B MLX MXFP4

An Apple Silicon MLX-VLM MXFP4 release of the original meta-models/Muse-Glimmer-30B model, developed by Meta Superintelligence Lab. This conversion is pinned to source revision 97c77dff50b2797bcc558fa2d909761dbc575c59 and is intended for local text-and-image inference with MLX.

The original model, this conversion, and the DFlash assistant are Apache-2.0. Read the upstream model card and USAGE_POLICY.md before deployment.

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 by the original model; not validated end-to-end here
Quantization	RTN mxfp4, 4-bit, group size 32
Quantization scope	Eligible language and bridge linear layers; vision_tower remains unquantized
Release size	Approximately 18.57 GB
Validated host	Apple M2 Max with 32 GB unified memory

Requirements

• macOS on Apple Silicon.
• Validated environment: macOS 26.5.1, Python 3.13.6, MLX 0.32.0; see reports/runtime.json for the complete package set.
• Approximately 19 GB for model files, plus memory for MLX, the KV cache, prompt processing, and the operating system.
• 32 GB unified memory was validated for autoregressive text and vision use.
• DFlash was not usable on the tested 32 GB host; see the experimental section below.
• The Hugging Face hf CLI and network access are needed only if downloading the separate DFlash assistant with the command below.

Installation and reproducibility

• MLX-VLM runtime: 132fc5dea46a085f39400cb101d1ad7c1dabdd49
• Muse DFlash support: experimental PR #1842
• Source revision: 97c77dff50b2797bcc558fa2d909761dbc575c59
• DFlash assistant: meta-models/Muse-Glimmer-30B-assistant at e8192f3a8f617f74be2ce220360c89ef4789f39f
• Exact tensor reference: 1857/1857 tensors match mlx-community/Muse-Glimmer-30B-mxfp4 at a50ab3a310fcc98a0273452e17a95a23d2e70e26

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

This exact MLX-VLM revision is required because Muse Glimmer support is not yet part of a stable release.

Usage

Text generation

mlx_vlm.generate \
  --model bsisduck/Muse-Glimmer-30B-MLX-MXFP4 \
  --prompt "Write a quicksort in Python." \
  --max-tokens 512 --temperature 1.0 --top-p 0.95 --top-k 64

Muse Glimmer's published recommended sampling is temperature 1.0, top-p 0.95, and top-k 64. The benchmark below uses greedy decoding because that is the published DFlash speed protocol and permits output equality checks.

Image understanding

mlx_vlm.generate \
  --model bsisduck/Muse-Glimmer-30B-MLX-MXFP4 \
  --image ./image.jpg \
  --prompt "Describe this image and identify any important details." \
  --max-tokens 512 --temperature 1.0 --top-p 0.95 --top-k 64

DFlash (experimental)

On the tested 32 GB M2 Max, the target and pinned assistant loaded together, but the 8-token DFlash warmup did not complete after 43.6 minutes. The process footprint reached approximately 22.5 GB while free system memory fell to 4%, with heavy compression and swap. No local DFlash throughput or output-equality result is claimed. DFlash likely requires more available memory, but no minimum-memory requirement or higher-memory MLX configuration was validated here. The 32 GB validation covers autoregressive generation only.

hf download meta-models/Muse-Glimmer-30B-assistant \
  --revision e8192f3a8f617f74be2ce220360c89ef4789f39f \
  --local-dir ./Muse-Glimmer-30B-assistant

mlx_vlm.generate \
  --model bsisduck/Muse-Glimmer-30B-MLX-MXFP4 \
  --draft-model ./Muse-Glimmer-30B-assistant \
  --draft-kind dflash \
  --prompt "Write a quicksort in Python." \
  --max-tokens 512 --temperature 0

The upstream model card reports 1.5x on M4 Max (23.7 to 37.8 tok/s) and 1.8x on M5 Max (26.6 to 50.2 tok/s). Those measurements used the upstream 17 GB K-Quant target with ExecuTorch, batch size 1, and greedy decoding. They are not MLX measurements and are not directly comparable to the local result below.

Validation

Check	Result
Serialized tensor comparison	1857/1857 exact
Text smoke test	Passed; 71 prompt tokens, 48 generated tokens, expected result 126 present
Vision smoke test	Passed; 64x64 red-square fixture, 79 prompt tokens, 32 generated tokens, identified as red
Quantization metadata	Passed; MXFP4, 4-bit, group size 32
Vision tower policy	Passed; the vision tower is unquantized

These checks establish artifact integrity, reproducibility, loadability, and basic text/vision operation. They are not a task-quality comparison against the BF16 source model. Exact agreement with another serialized MXFP4 release does not prove quality equivalence to BF16.

The documented 131,072+ context limit was not exercised on the 32 GB validation host. Long prompts and larger or more complex images may require more memory.

Local throughput

Measured on the host described in reports/benchmark.json, batch size 1, four prompt categories, greedy decoding, and an 8-token warmup:

Mode	Mean tok/s	Median tok/s	Peak MLX memory
Autoregressive	2.73	2.37	18.86 GB
DFlash	Not completed on 32 GB	Not completed on 32 GB	~22.5 GB process footprint

The baseline completed all four prompts. DFlash loaded but did not finish its 8-token warmup under severe unified-memory pressure, so speedup and greedy-output equality were not measured locally.

The complete protocol, per-prompt results, environment, and package versions are available in reports/benchmark.json.

DFlash numerical caveat

The MLX-VLM PR author reports that a real 4-bit Muse target can choose different greedy tokens between token-at-a-time and block verification because quantized matrix multiplication is shape-dependent. Target verification still occurs, but do not assume bit-identical output on every prompt. No local equality result is claimed because DFlash did not complete warmup on this 32 GB host.

Intended use

This release is intended for local research and application development on Apple Silicon, including:

• text generation and reasoning;
• coding and tool-using agents with application-level safeguards;
• image and document understanding;
• evaluation of MLX-VLM integration and MXFP4 deployment.

Audio input/output is not supported. Uses that violate applicable law, the Apache-2.0 license, or the included usage policy are out of scope.

Limitations and safety

• The model may produce inaccurate, biased, unsafe, or objectionable output.
• Quantization can change outputs and may reduce quality in edge cases. No full BF16-versus-MXFP4 quality benchmark was run for this release.
• Performance can vary substantially with prompt length, memory pressure, OS, and MLX/runtime versions.
• DFlash depends on an experimental runtime revision and was not validated as usable on the tested 32 GB host.
• Agentic deployments should validate tool calls, require confirmation for irreversible actions, and apply safeguards appropriate to the use case.
• The original model was trained on more than 100 languages, but performance was not validated here across that language set.

The original model's safety evaluations and broader limitations are documented in its model card and methodology report.

Repository scope

This repository publishes one model variant: the MXFP4 target described above. It is not a multi-variant repository and intentionally does not duplicate Meta's 5.1 GB BF16 assistant. Download the pinned official assistant separately with the hf download command shown above before using DFlash.

Release evidence is included in:

• reports/conversion.json — source and conversion settings;
• reports/runtime.json — pinned package versions;
• reports/smoke.json — text and vision smoke checks;
• reports/tensor_audit.json — exact tensor comparison;
• reports/benchmark.json — throughput and DFlash status.