Yes, just — 0.9 GB spare
Muse Glimmer 30B at Q4_K_M fits your Radeon RX 7900 XT entirely on the GPU at 8K context, at an estimated 29 tokens per second. Past 76K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Meta's first open weights since Llama 4: a dense 30B distilled from Muse Spark for always-on local agents. Two KV heads keep the cache small.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 29.5 GB | 30.3 GB | — | ~2.8 | −0.1% ppl | 11.9 GB over |
| Q6_K | 22.8 GB | 23.5 GB | — | ~5.6 | −0.4% ppl | 5.1 GB over |
| Q5_K_M | 19.7 GB | 20.4 GB | — | ~11 | −0.8% ppl | 2.0 GB over |
| Q4_K_M | 16.8 GB | 17.5 GB | 76K | 29 | −1.9% ppl | Recommended |
| Q3_K_M | 13.6 GB | 14.3 GB | 128K | 36 | −5.4% ppl | Long context |
| Q2_K | 11.6 GB | 12.4 GB | 128K | 42 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. This model interleaves sliding-window layers (2048 tokens, 13 of 52 layers global), which is why its cache barely grows with context.
How to run it
$ llama-server \
-hf meta-models/Muse-Glimmer-30B:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.