readygpu

Muse Glimmer 30B

29.8B parameters · 128k context · Apache 2.0 · released 2026-08-09

chatcodingagenticvisionreasoning

In short

Muse Glimmer 30B is an open model from meta-models with 29.8B parameters. At Q4_K_M, Muse Glimmer 30B needs about 17 GB of memory at 8k context and 19 GB at 128k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in Canada that runs Muse Glimmer 30B fully in memory at 128k context is the AMD Radeon RX 7900 XTX, $1,800 at eBay, at about 24 tok/s (estimated).

Updated . Speeds are estimates; how they are worked out.

What to run it on in Canada

Fully in memory at 128k context, priced at retailers today.

All hardware that runs it

Everything we price, best value first. Two cards split the model by layers.

HardwareQuantFitsDecodePricetok/s per 1,000
AMD Radeon RX 7900 XTXQ5_K_Mfits24 tok/s$1,80013.4
Intel Arc Pro B60Q5_K_Mfits10 tok/s$1,9325.4
AMD Radeon AI PRO R9700Q6_Kfits14 tok/s$2,7855.0
NVIDIA GeForce RTX 3090Q5_K_Mfits27 tok/s$2,9009.1
NVIDIA RTX PRO 6000 BlackwellQ8_0fits33 tok/s$3,22510.1
NVIDIA GeForce RTX 4090Q5_K_Mfits29 tok/s$3,3008.6
NVIDIA GeForce RTX 3090 TiQ5_K_Mfits29 tok/s$3,9997.1
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits4 tok/s$5,0990.8
ASUS Ascent GX10 128GBQ8_0fits5 tok/s$8,2990.6
NVIDIA GeForce RTX 5090Q6_Kfits44 tok/s$10,0004.4
NVIDIA DGX Spark 128GBQ8_0fits5 tok/s$11,0140.5
Mac Studio M3 Ultra 512GBQ8_0fits16 tok/s$28,0000.6
Mac Studio M4 Max 128GBQ8_0fits10 tok/s––
Mac mini M4 Pro 64GBQ8_0fits5 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits4 tok/s––
Mac Studio M3 Ultra 256GBQ8_0fits16 tok/s––
Mac mini M4 Pro 48GBQ8_0fits5 tok/s––
Mac Studio M4 Max 64GBQ8_0fits10 tok/s––
Mac Studio M3 Ultra 96GBQ8_0fits16 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits4 tok/s––
Mac mini M4 32GBQ4_K_Mfits4 tok/s––
NVIDIA Tesla P40Q5_K_Mfits10 tok/s––
Beelink GTR9 Pro Ryzen AI Max+ 395 128GBQ8_0fits4 tok/s––
Intel Arc B580Q4_K_Moffload5 tok/s$5069.5
NVIDIA GeForce RTX 3060 12GBQ4_K_Moffload5 tok/s$5309.2
NVIDIA GeForce RTX 3080 10GBQ4_K_Moffload5 tok/s$7696.0
AMD Radeon RX 7800 XTQ4_K_Moffload8 tok/s$8009.9
AMD Radeon RX 9060 XT 16GBQ4_K_Moffload6 tok/s$9006.7
NVIDIA GeForce RTX 5060 Ti 16GBQ4_K_Moffload8 tok/s$1,1206.8
AMD Radeon RX 9070 XTQ4_K_Moffload8 tok/s$1,2006.7
AMD Radeon RX 7900 XTQ4_K_Moffload22 tok/s$1,20017.9
NVIDIA GeForce RTX 5070Q4_K_Moffload6 tok/s$1,2594.4
Intel Arc A770 16GBQ4_K_Moffload7 tok/s$1,3035.7
NVIDIA GeForce RTX 4070 SuperQ4_K_Moffload5 tok/s$1,4003.7
AMD Radeon RX 7600 XTQ4_K_Moffload6 tok/s$1,4194.0
NVIDIA GeForce RTX 4060 Ti 16GBQ4_K_Moffload6 tok/s$1,4404.3
NVIDIA GeForce RTX 4070 Ti SuperQ4_K_Moffload9 tok/s$1,5305.8
NVIDIA GeForce RTX 5070 TiQ4_K_Moffload10 tok/s$1,8005.3
NVIDIA GeForce RTX 4080 SuperQ4_K_Moffload9 tok/s$1,9004.8
NVIDIA GeForce RTX 5080Q4_K_Moffload10 tok/s$2,5703.8

1 other card or machine we price cannot hold it at this context.

Memory needed

Weights, KV cache and runtime buffers at each context length (KV cache at fp16).

QuantFile@8k@32k@64k@128k
Q4_K_Mgguf16.1 GB17.0 GB17.5 GB18.1 GB19.3 GB
mlx-4bitmlx18.1 GB19.0 GB19.5 GB20.0 GB21.2 GB
Q5_K_Mgguf18.7 GB19.6 GB20.1 GB20.7 GB21.9 GB
mlx-5bitmlx21.3 GB22.3 GB22.7 GB23.3 GB24.5 GB
Q6_Kgguf21.8 GB22.7 GB23.2 GB23.8 GB25.0 GB
mlx-6bitmlx24.6 GB25.5 GB26.0 GB26.5 GB27.7 GB
Q8_0gguf30.1 GB31.0 GB31.5 GB32.1 GB33.3 GB
mlx-8bitmlx31.1 GB32.0 GB32.5 GB33.0 GB34.2 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/bartowski/Muse-Glimmer-30B-GGUF:Q4_K_M
    • LM Studiolms get bartowski/Muse-Glimmer-30B-GGUF@q4_k_m
    • llama.cppllama-server -hf bartowski/Muse-Glimmer-30B-GGUF:Q4_K_M
  • mlx-4bit
    • LM Studiolms get mlx-community/Muse-Glimmer-30B-4bit
    • MLXmlx_lm.server --model mlx-community/Muse-Glimmer-30B-4bit
  • Q5_K_M
    • Ollamaollama run hf.co/bartowski/Muse-Glimmer-30B-GGUF:Q5_K_M
    • LM Studiolms get bartowski/Muse-Glimmer-30B-GGUF@q5_k_m
    • llama.cppllama-server -hf bartowski/Muse-Glimmer-30B-GGUF:Q5_K_M
  • mlx-5bit
    • LM Studiolms get mlx-community/Muse-Glimmer-30B-5bit
    • MLXmlx_lm.server --model mlx-community/Muse-Glimmer-30B-5bit
  • Q6_K
    • Ollamaollama run hf.co/bartowski/Muse-Glimmer-30B-GGUF:Q6_K
    • LM Studiolms get bartowski/Muse-Glimmer-30B-GGUF@q6_k
    • llama.cppllama-server -hf bartowski/Muse-Glimmer-30B-GGUF:Q6_K
  • mlx-6bit
    • LM Studiolms get mlx-community/Muse-Glimmer-30B-6bit
    • MLXmlx_lm.server --model mlx-community/Muse-Glimmer-30B-6bit
  • Q8_0
    • Ollamaollama run hf.co/bartowski/Muse-Glimmer-30B-GGUF:Q8_0
    • LM Studiolms get bartowski/Muse-Glimmer-30B-GGUF@q8_0
    • llama.cppllama-server -hf bartowski/Muse-Glimmer-30B-GGUF:Q8_0
  • mlx-8bit
    • LM Studiolms get mlx-community/Muse-Glimmer-30B-8bit
    • MLXmlx_lm.server --model mlx-community/Muse-Glimmer-30B-8bit

Notes

Dense 30B model with a ~1.8B vision encoder; 4-bit quantized variants target 24/32 GB VRAM (full precision ~64GB). Ships with a DFlash speculative decoding drafter.

Questions

How much VRAM does Muse Glimmer 30B need?
At Q4_K_M, Muse Glimmer 30B needs about 17 GB of memory at 8k context and 19 GB at 128k, counting the weights, the KV cache and runtime buffers.
What is the best GPU for running Muse Glimmer 30B locally?
The cheapest hardware in Canada that runs Muse Glimmer 30B fully in memory at 128k context is the AMD Radeon RX 7900 XTX, $1,800 at eBay, at about 24 tok/s (estimated). The fastest we price is the NVIDIA GeForce RTX 5090, at about 44 tok/s.
Can I run Muse Glimmer 30B on a 16 GB GPU?
Only partly: at 128k context it does not fit in 16 GB, so part of it runs from system RAM and it slows to about 8 tok/s on the AMD Radeon RX 7800 XT (estimated).
Can I run Muse Glimmer 30B on a 24 GB GPU?
Yes: at Q5_K_M it fits fully in 24 GB at 128k context. On the NVIDIA GeForce RTX 4090 it runs at about 29 tok/s (estimated).
How do I run Muse Glimmer 30B locally?
With Ollama: ollama run hf.co/bartowski/Muse-Glimmer-30B-GGUF:Q4_K_M (the Q4_K_M build). LM Studio and llama.cpp commands are on this page.
What licence is Muse Glimmer 30B released under?
Apache 2.0, which allows commercial use.

More hardware for local AI