How much VRAM does Muse Glimmer 30B need at Q4_K_M? Native 131K context

By TechCompare · Updated · How we check

Muse Glimmer 30B at Q4_K_M needs about 22.2 GB at its native 131K context, so it fits a 24 GB card. The calculator also puts an extended 256K run near 24.1 GB, but 256K should not be described as native until the serving stack and model release document that support.

Calculator

Estimated VRAM required

22.2 GB

30B params at Q4_K_M, 131,072 token context, batch 1, inference.

Weights
18.3 GB
KV cache
1.9 GB
Overhead
2.1 GB

Estimate accuracy: Weights within ~3% of most GGUF files from Q4_K_M up, ~7% for Q3_K_M and Q2_K. KV cache within ~5% for standard GQA models, ~10% for MLA or hybrid state estimates. Real VRAM may vary with framework (vLLM vs llama.cpp vs Transformers), Flash Attention, and driver overhead.

Sliding-window attention applied: This model caps 3 of every 4 layers at a 2048-token window. KV cache estimate is 74% smaller than naive full-attention math at this context length.

Hardware that fits

Unified memory machines show a model budget, not their installed memory. We hold back 8 GB for the system, or use the vendor's GPU limit when that's lower. It's a planning estimate, not a hard limit.

RTX 3090
Consumer
24 GB
93% used
Apple M5 32GB
Unified
~24 GB
93% of estimate
RTX 5090
Consumer
32 GB
70% used
A100 40GB
Datacenter
40 GB
56% used

Muse Glimmer 30B at Q4_K_M needs about 22.2 GB of VRAM at its native 131K context: 18.3 GB of weights, 1.9 GB of KV cache, and 2.1 GB of overhead.

The official architecture has 52 layers with three local attention layers for every global layer, so the 256K calculator scenario is an extended run rather than native support. Apache 2.0 licensing and a sub-24 GB native estimate make it a practical single-GPU model.

How this is calculated

The released config uses 52 layers with a [Local, Local, Local, Global] pattern. Three out of every four layers cap their KV cache at a 2048-token sliding window, while 13 global layers retain the full sequence. With 2 KV heads and head_dim 128, that gives about 1.83 GB of FP16 KV cache at native 131,072 tokens and a calculated total of 22.2 GB.

At an extended 262,144-token scenario, the hybrid cache is about 3.58 GB and the total is about 24.1 GB. Using full-context KV storage for all 52 layers would be about 6.98 GB at native 131K. The 30B total at Q4_K_M is 18.3 GB of weights, and Apache 2.0 licensing means no commercial strings.

Verdict

Run Glimmer at its native 131K, where the estimate is 22.2 GB and an RTX 3090 fits it. The cache is 1.83 GB of FP16 from 13 global layers plus 39 local layers.

The calculator's 256K scenario gives 24.1 GB with a 3.58 GB cache, but 131K is the released native position limit in the official config. Against Gemma 4 31B dense at a similar footprint, the main difference is the license: Apache 2.0 for Glimmer, a Google license for Gemma.

More Meta scenarios

Kimi K3.1 Pro at Q4_K_M
A 2.78T expert pool with 24 latent attention layers and fixed state in the other 69.
View details
Mistral Large 4 at Q4_K_M
A 1T expert pool with 63 latent cache owners and 44 layers that also store an index key.
View details
Muse Spark Open at Q4_K_M
Only 16 of its 62 layers grow with context, and the other 46 hold a 2,048 token window.
View details

Frequently asked questions

Is Muse Glimmer the successor to Llama 4?
It's Meta's 2026 open release rather than a numbered Llama sequel. Glimmer is multimodal (28B text + 2B vision), Apache 2.0, and built around hybrid sliding-window attention. There's no confirmed Llama 5 open-weight release as of this writing.
Muse Glimmer 30B or Qwen 3.8 27B?
Glimmer if you want Apache 2.0 licensing and a smaller native footprint (about 22.2 GB at 131K, or about 24.1 GB in the calculator's extended 256K scenario). Qwen 3.8 27B if you need its longer native window and can budget the much larger long-context footprint. Pick on the context limit, license, and ecosystem you actually need.