TechCompare LogoTechCompare

How much VRAM does Muse Glimmer 30B need at Q4_K_M? Meta's open-weights return

Muse Glimmer 30B at Q4_K_M fits one 32 GB card at its full 256K context (28 GB total), and one 24 GB card if you cap context at 128K. It's the pick when you want Meta-quality weights with Llama-clean licensing in a single-GPU footprint.

Muse Glimmer 30B at Q4_K_M needs about 28 GB of VRAM at its native 256K context - 17 GB of weights plus an 8.7 GB cache kept small by hybrid sliding-window attention plus 2.6 GB overhead. It's Meta's first major open release since the Llama 4 line, Apache 2.0 licensed, and sized for a single GPU.

By TechCompare · Updated

Total VRAM required
28.1 GB
Muse Glimmer 30B at Q4_K_M
Weights
16.8 GB
30B params
KV cache
8.7 GB
256K tokens, FP16 KV

Calculator

Estimated VRAM required

28.1 GB

30B params at Q4_K_M, 262,144 token context, batch 1, inference.

Weights
16.8 GB
KV cache
8.7 GB
Overhead
2.6 GB

Estimate accuracy: Weights within ~2%. KV cache within ~5% for standard GQA models, ~10% for MLA (DeepSeek). Real VRAM may vary with framework (vLLM vs llama.cpp vs Transformers), Flash Attention, and driver overhead.

Sliding-window attention applied: This model caps 4 of every 5 layers at a 1024-token window. KV cache estimate is 80% smaller than naive full-attention math at this context length.

Hardware that fits

RTX 5090
Consumer
32 GB
88% used
A100 40GB
Datacenter
40 GB
70% used
Apple M3 Max 64GB
Unified
48 GB
58% used
RTX 6000 Ada
Pro
48 GB
58% used

Just barely too small

RTX 3090
Consumer
24 GB
short by 4.1 GB

How this is calculated

Glimmer alternates full-attention and 1024-token sliding-window layers, with full attention every 5th layer. That's the Gemma-style recipe that keeps the KV cache at 8.7 GB at 256K instead of the ~100 GB a plain full-attention 40-layer model would burn at the same window. The 30B total (28B text + 2B vision) at Q4_K_M is 16.8 GB of weights, and Apache 2.0 licensing means no commercial strings.

Verdict

The 28 GB resident total is 16.8 GB of Q4 weights, 8.7 GB of FP16 KV cache from the hybrid SWA recipe (full attention every 5th layer, 1024-token window elsewhere), and 2.6 GB of overhead. A 32 GB card or a 4090 with 128K context cap runs it cleanly. The comparison that matters: Gemma 4 31B dense at the same class footprint but with a Google license, versus Glimmer with Apache 2.0, and Qwen 3.8 27B with a longer native window but a cache that punishes long contexts.

More Meta scenarios

DeepSeek V4 Pro 1.6T (MoE) at Q4_K_M
DeepSeek V4 Pro 1.6T at Q4_K_M with the full 1M-token context needs about 1012 GB of VRAM with every expert resident - this is the real shape of the model and the number to plan a deployment against.
View details ➜
Llama 4 Scout (17B/109B) at Q4_K_M
Llama 4 Scout at Q4_K_M with its native 10M context needs about 2231 GB of VRAM with all 109B params resident - that's the number you size hardware against.
View details ➜
gpt-oss 20B (MoE) at Q4_K_M
gpt-oss 20B at Q4_K_M with native 128K context needs about 19.4 GB of VRAM with all experts resident, dropping to roughly 9.3 GB with active-only weight loading.
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 cache that stays small at long context (28 GB at 256K). Qwen 3.8 27B if you need the full 1M native window and can budget the 244 GB it costs at that length. At 128K both fit a single GPU; pick on license and ecosystem.