Muse Glimmer 30B — Q2_K

A low-memory Q2_K quantization of Meta’s Muse Glimmer 30B, built directly from the official BF16 weights.

Designed to make Muse Glimmer 30B practical on lower-memory hardware, including Apple Silicon Macs.

Quick Start

ollama run yogeshjog/muse-glimmer-q2

Or explicitly use the Q2_K tag:

ollama run yogeshjog/muse-glimmer-q2:q2_k

Model

The model was quantized directly from the official BF16 weights.

It was not requantized from an existing Q4 or other quantized checkpoint.

Official Muse-Glimmer-30B BF16
              │
              ▼
          BF16 GGUF
              │
              ▼
       llama.cpp Q2_K
              │
              ▼
       9.95 GiB GGUF
              │
              ▼
            Ollama

Tested on Apple Silicon

Created and tested locally on:

MacBook Pro
Apple M3 Max
36 GB unified memory

Tested with Ollama at a configured 32K context window:

NAME                      SIZE     PROCESSOR    CONTEXT
muse-glimmer-q2:latest    11 GB    100% GPU     32768

The model successfully loaded with 100% GPU offload.

Actual memory consumption can vary depending on context utilization, KV cache configuration, Ollama version, and workload.

Why Q2_K?

Muse Glimmer 30B is relatively large for consumer hardware.

Q2_K reduces the model weights to approximately:

9.95 GiB
3.07 BPW

The objective is simple:

How much of Muse Glimmer 30B can we retain while bringing the model down to roughly 10 GiB?

This makes it particularly interesting for:

• Apple Silicon
• lower-memory GPUs
• laptops and workstations
• local inference
• larger context experiments
• memory-constrained environments

Quality Warning

Q2_K is an aggressive quantization.

Expect some quality degradation compared with BF16 and higher-bit quantizations such as Q4, Q5, Q6, or Q8.

Potentially affected capabilities include:

• reasoning
• coding
• instruction following
• factual accuracy
• structured output
• long-context performance
• multimodal performance

Evaluate the model for your particular workload.

Quantization Details

Observed llama.cpp output:

model size = 53131.48 MiB (16.00 BPW)
quant size = 10179.12 MiB (3.07 BPW)

quantize time = 79786.76 ms

The entire conversion and quantization process was performed locally on an M3 Max with 36 GB unified memory.

No cloud GPU was required.

GGUF Download

The raw Q2_K GGUF and full model card are available on Hugging Face:

https://huggingface.co/yogeshjog/Muse-Glimmer-30B-Q2_K-GGUF

Reproduce the Quantization

The quantization was produced using:

https://github.com/phygineer/quantize

Pipeline:

Hugging Face BF16
      ↓
BF16 GGUF
      ↓
llama.cpp
      ↓
Q2_K

Original Model

Original model:

https://huggingface.co/meta-models/Muse-Glimmer-30B

Please refer to the original model repository for authoritative documentation, intended use, limitations, safety information, and licensing details.

Credits

• Original model: Meta Superintelligence Labs
• Base model: meta-models/Muse-Glimmer-30B
• Quantization: Q2_K
• GGUF / quantization tooling: llama.cpp
• Quantization pipeline: phygineer/quantize
• Quantized by: yogeshjog

Disclaimer

This is an unofficial community quantization and is not affiliated with or endorsed by Meta.

Low-bit quantization changes model weights and can materially affect model behavior and output quality.