16GB
Official publisher quants do not fit with a practical working budget. Community 2-bit experiments are not the same baseline.
Local compatibility answer · Updated Aug 11, 2026
Can I run Muse Glimmer 30B locally?
Yes on 24GB-class hardware—but 32GB is the comfortable tier. Meta's compact 16.8GB model is built for a 24GB memory envelope. AICanRun estimates a 20.5GB base working set at 4K context; vision and DFlash companions can bring the planning total close to 23.5GB.
SHORT ANSWER24GB: tight yes32GB+: recommended45 of 54 tracked Macs memory-fit the official compact quant.
24GB
K-Quant-17GB is Meta's target. Fine for the base model at short context; the complete vision + DFlash stack leaves little margin.
32GB
The sensible Mac and PC tier. The dynamic quant's file-based planning total is ~26.6GB, leaving useful margin.
64GB+
Comfortable for long sessions and larger contexts. Meta lists 64GB VRAM as the target for full precision.
The number that matters
Why 17GB does not mean you only need 17GB
File size is only the starting point. The loaded weights, KV cache, runtime, vision encoder, and optional DFlash drafter share the same memory envelope.
weightsKVappvision + DFlash
17.6GB loaded weights2.1GB 4K KV cache estimate0.8GB runtime allowance~3GB companion files
The 20.5GB base figure is an AICanRun planning estimate, not a measured peak allocation. Companion file sizes are not guaranteed to equal resident memory, and longer context uses more.
| Number you may see | What it actually describes | How to use it |
|---|---|---|
| ~14GB | Experimental community 2-bit build | Possible loading experiment, not Meta's quality baseline. |
| 16.8GB | Official K-Quant-17GB main GGUF | Download size—not the full working set. |
| ~20.5GB | Main model + 4K KV + runtime estimate | Useful baseline for text-only planning. |
| ~23.5GB | Baseline plus 1.4GB vision projector and 1.6GB DFlash files | Explains why a 24GB setup is tight. |
| 24 / 32 / 64GB | Meta's target hardware tiers | Compact quant / dynamic quant / full precision. |
Exact Apple configurations
Can my Mac run Muse Glimmer?
Fixed to K-Quant-17GB and 4K context. Speed is a bandwidth-based range, not a measured benchmark.
| Mac | Installed memory | Working budget | Headroom | Estimated decode | Answer |
|---|---|---|---|---|---|
| Mac mini M4 | 16GB | 13GB | -7.5GB | — | NO |
| Mac mini M4 | 24GB | 21GB | +0.5GB | ~3.6–5 tok/s | TIGHT |
| Mac mini M4 | 32GB | 28.8GB | +8.3GB | ~3.6–5 tok/s | YES |
| MacBook Air M4 | 24GB | 21GB | +0.5GB | ~3–4.1 tok/s | TIGHT |
| MacBook Pro M4 Pro | 24GB | 21GB | +0.5GB | ~8.1–11.4 tok/s | TIGHT |
| MacBook Pro M4 Max | 36GB | 32.4GB | +11.9GB | ~12.2–17.1 tok/s | YES |
| MacBook Pro M5 Pro | 48GB | 43.2GB | +22.7GB | ~9.1–12.8 tok/s | YES |
| MacBook Pro M5 Max | 64GB | 57.6GB | +37.1GB | ~18.3–25.6 tok/s | YES |
How to run Muse Glimmer locally
Choose the runtime that matches your machine
Support landed quickly after release. Update your runtime before treating a model-load error as a bad download.
Ollama with MLX
Ollama publishes an Apple-silicon MLX package with image and DFlash support.
ollama run muse-glimmer:30b-mlxLM Studio Bionic
Install LM Studio Bionic, open Settings → Explore, find Muse Glimmer, choose a quant that matches your memory, and download.
Official LM Studio guide ↗llama.cpp + official GGUF
Use a current llama.cpp build. The main GGUF runs text; add mmproj-kquant.gguf for images and the DFlash drafter when you want speculative decoding.
Safe first run: start at 4K or 8K context, confirm text generation, then add vision or DFlash and increase context one step at a time. Compare DFlash on and off for your own workload—speculative decoding can lose performance when draft acceptance is low. A model loading successfully does not mean a 128K context will fit.
Early field notes
What real users are finding so far
These are directional patterns from day-one community use, not AICanRun benchmarks. Treat them as debugging clues rather than promised performance.
Acceleration is workload-dependent
Repetitive coding edits can produce high draft acceptance and dramatic gains. Other prompts and some backends show low acceptance or even slower output. Always A/B the same task before keeping it enabled.
Better in a loop than a one-shot
Early users often like its tool-call discipline and recovery, while large one-shot coding tasks receive mixed reviews. Give it tools, tests, and permission to iterate instead of judging it only on a single giant generation.
Watch token use and stopping behavior
Some runs spend heavily on reasoning or yield control early. Use an explicit reasoning level, leave enough output budget, and test whether a lower reasoning mode is more productive for routine edits.
Closest local alternatives
Muse Glimmer vs Qwen 3.6 27B and Gemma 4 31B
Meta's benchmark claims are vendor-reported. This table focuses on the hardware and product decision.
| Model | Comparable local file | Local-hardware angle | Best reason to choose it |
|---|---|---|---|
| Muse Glimmer 30B | 16.8GB K-Quant | Official 24GB full-stack target with optional DFlash acceleration. | Tool use, recovery, multimodal local agents. |
| Qwen 3.6 27B | 16.8GB Q4_K_M | Similar weight footprint with many quant choices. | A mature general/coding alternative and useful comparison baseline. |
| Gemma 4 31B | 18.3GB Q4_K_M | Slightly larger Q4 file in the current catalog. | Another dense 30B-class choice for quality-first local use. |
Direct answers
Muse Glimmer FAQ
Can I run Muse Glimmer 30B locally?
Yes, if you have a 24GB-class GPU or enough Apple unified memory. Meta targets 24GB with K-Quant-17GB, 32GB with K-Quant-Dynamic, and 64GB for full precision. AICanRun estimates a 20.5GB base working set at 4K context before the optional vision and DFlash companion files.
Can a 16GB GPU or 16GB Mac run Muse Glimmer?
Not with Meta's two official publisher GGUF weights at a practical full-offload setting. Experimental community 2-bit builds may load in less memory, but they are a different quality and support tradeoff and are not the baseline used on this page.
Can a 24GB Mac run Muse Glimmer?
The official 17GB main model fits our conservative 4K text baseline at about 20.5GB, but a 24GB Mac has very little working headroom after macOS. Loading the vision projector and DFlash drafter can push the file-based planning total to roughly 23.5GB, so 32GB unified memory is the safer recommendation.
Why do Muse Glimmer memory recommendations range from 14GB to 32GB?
Those numbers describe different things: experimental 2-bit community quants, the main weight download, the loaded model plus KV cache and runtime, or a complete vision and speculative-decoding stack. They should not be treated as interchangeable minimum requirements.
How do I run Muse Glimmer locally?
Apple-silicon users can use Ollama's MLX package; Mac and PC users can use LM Studio Bionic; advanced users can load Meta's official GGUF with a current llama.cpp build. Download first, start with a 4K or 8K context, then raise context only after checking memory headroom.
Can a phone run Muse Glimmer 30B?
Not at Meta's official publisher quants in our current device database. 0 of the phones we track pass the conservative memory check, so Muse Glimmer should be treated as a desktop or high-memory Mac model.
Meta release and hardware targets · Meta model card · Official GGUF repository
Hardware-fit and speed figures without a ✓ mark are AICanRun estimates. Meta performance and quality claims remain vendor-reported.