Can I Run Muse Glimmer 30B Locally? (2026): VRAM by Quant

Bottom Line

• 24GB is the sweet spot. Q4_K_M (16.12 GiB) plus the DFlash drafter plus a 131K context is about 19 GiB. Add the vision file at 32K and it is still under 23 GiB.
• 16GB works at 3 bits. Q3_K_M (13.00 GiB), text only, 32K context.
• 32GB runs Q6_K with everything. Vision, DFlash and a 131K context, about 29 GiB.
• The cache is tiny. 39 of 52 layers look back only 2,048 tokens. A full 131K context costs about 1.7 GiB.
• DFlash is the reason to pick it. Meta measured 233.4 tok/s on an RTX 5090 with DFlash against 74.9 without.
• Apache 2.0. No commercial limits, unlike Meta’s big Muse Spark Open.

Ready to buy? See the tested hardware list with current prices.

What Muse Glimmer 30B Is

Spec	Value
Released	10 August 2026 (Hugging Face blog)
Parameters	about 29.6B, dense: a 2B vision encoder plus a 28B text decoder
Layers	52: 39 sliding-window (2,048 tokens), 13 full attention
Attention	32 query heads, 2 key-value heads, head size 128
Context	131,072 in config.json (see below)
Input	Text, images, video frames
Licence	Apache 2.0

The sources disagree on context length. config.json says 131,072. The model card says “131,072 or more”. Unsloth says the default is 131,072 and it goes up to 262,144. Ollama lists 128K. Plan on 131K.

It is built for agents. Meta’s card names OpenClaw and Hermes Agent, and lists four reasoning levels: low, medium, high and xhigh. Our OpenClaw setup guide shows how to connect it.

Every GGUF Size, in Bytes

From bartowski’s repo, summed through the Hugging Face API. GiB is what your VRAM counts.

Quant	GB	GiB	Fits
IQ2_XXS	8.92	8.31	12GB, low quality
IQ2_M	10.66	9.93	12GB
Q2_K	11.04	10.28	12GB, tight
IQ3_M	13.11	12.21	16GB
Q3_K_M	13.96	13.00	16GB
IQ4_XS	15.44	14.38	16GB, very tight
Q4_K_S	16.32	15.20	24GB
Q4_K_M	17.31	16.12	24GB
Q5_K_M	20.11	18.72	24GB, text
Q6_K	23.41	21.81	32GB
Q8_0	29.61	27.58	32GB text, or 48GB Mac
BF16	55.73	51.90	96GB+ Mac, 128GB box

Add-ons:

• Vision file (mmproj f16): 3.58 GiB. Skip it for text-only agents.
• DFlash drafter Q4_0: 1.35 GiB. Q8_0: 2.54 GiB.

Why the Cache Costs So Little

We read these values from config.json: 2 key-value heads, head size 128, 13 full-attention layers.

• Full-attention layers: 13 x 2 heads x 128 x 2 (key and value) x 2 bytes = 13 KiB per token at 16-bit.
• Sliding layers: a fixed 78 MiB.

Context	Cache (16-bit)
32,768	0.5 GiB
131,072	1.7 GiB
262,144	3.3 GiB

This is why a long context is nearly free. Most 30B models need several GiB for 131K. Glimmer needs less than 2 GiB. It is our estimate; check the cache line in your llama-server log.

Best Quant by Hardware

GPU rows leave about 1 GiB for the driver and add a 0.5 GiB compute buffer. Mac rows use the default GPU limit, about 2/3 of unified memory on small Macs and about 3/4 on larger ones.

Hardware	Best setup	Total	Context
12GB GPU	IQ2_M, text only	~11 GiB	32K
16GB GPU	Q3_K_M, text only	~14 GiB	32K
24GB GPU	Q4_K_M + DFlash, text	~19.7 GiB	131K
24GB GPU, with images	Q4_K_M + DFlash + vision	~22 GiB	32K
32GB GPU	Q6_K + DFlash + vision	~29 GiB	131K
24GB Mac	Q3_K_M, text	~14 GiB	32K
32GB Mac	Ollama 30b-mlx (19GB)	~20 GiB	32K
48GB Mac	Q8_0 + vision	~33 GiB	131K
64GB Mac	Q8_0 + DFlash + vision	~35 GiB	131K
96GB+ Mac or 128GB box	BF16	~54 GiB	131K

On a 24GB card, choose between images and Q5. Q5_K_M (18.72 GiB) with a 131K context is about 21 GiB, text only. If your agent reads screenshots, stay at Q4_K_M and add the vision file.

Meta’s own targets agree. The model card lists a 17GB quant for 24GB of VRAM (about 1.0% degradation), a dynamic quant for 32GB (about 0.2%), and full precision for 64GB.

How Fast Is It?

Meta’s model card, batch size 1:

Hardware	Without DFlash	With DFlash	Speedup
RTX 5090	74.9 tok/s	233.4 tok/s	3.1x
M4 Max	23.7 tok/s	37.8 tok/s	1.5x
M5 Max	26.6 tok/s	50.2 tok/s	1.8x

DFlash is speculative decoding. A small drafter guesses several tokens and the main model checks them in one pass. The Hugging Face blog runs it with --spec-type draft-dflash --spec-draft-n-max 15. On Ollama, the 30b-mlx tag includes it.

The gain is biggest on NVIDIA. The RTX 5090 has spare compute for the check pass. Macs are closer to their limit, so they gain less.

What to Buy

For most people: a 24GB card. A used RTX 3090 24GB runs Q4_K_M with DFlash and a 131K context. An RTX 4090 runs the same files faster.

For Q6 with images and DFlash: 32GB. The RTX 5090 is the card Meta measured at 233.4 tok/s. The Radeon AI PRO R9700 32GB is the cheaper 32GB card, but DFlash support on AMD is not documented.

On a budget: 16GB at 3 bits. An RTX 5060 Ti 16GB runs Q3_K_M, text only. It works, but a 24GB card is the better buy for this model.

On a Mac: 32GB or more. A 48GB+ MacBook Pro runs Q8_0 with images. 24GB Macs are limited to 3 bits.

Honest Caveats

1. We did not run the model. Fit rows are arithmetic from file sizes and config.json.
2. The cache figure is our estimate. It ignores the vision tokens of an image prompt.
3. The 0.5 GiB compute buffer is a budget. We did not measure it.
4. The speeds are Meta’s. We have not seen independent DFlash numbers.
5. The Mac GPU limit is approximate. You can raise it with sysctl, but leave macOS room.

Sources

• Hugging Face blog: Muse Glimmer: architecture, release date, licence, DFlash flags, agent frameworks
• Muse Glimmer 30B model card: parameters, context, memory targets, DFlash speed table, sampling, reasoning levels
• config.json: layers, layer types, heads, head size, sliding window, context
• bartowski GGUF file tree: byte sizes, read 31 October 2026
• Unsloth Muse Glimmer guide: context default and maximum
• Ollama library: muse-glimmer: tags, sizes, 128K context, MLX tag with DFlash

See Also

• Muse Glimmer + OpenClaw: run Meta’s agent model offline: the setup steps
• Qwen3.8-27B vs Muse Glimmer 30B: which one to pick for a 24GB card
• Can I run Muse Spark Open locally?: Meta’s 816B model, 256GB minimum
• Best local LLM for 24GB VRAM: the other models for this tier
• Is 24GB VRAM enough?: when to step up to 32GB
• Is a used RTX 3090 worth it?: the cheapest 24GB card