---

license: apache-2.0

base_model: meta-models/Muse-Glimmer-30B

base_model_relation: quantized

tags: [gguf, llama.cpp, rocm, gfx1151, strix-halo, amd, ryzen-ai-max-395, rocmfpx, muse, glimmer, vision, multimodal]

language: [en]

pipeline_tag: image-text-to-text

---

> ### ⚠️ STOCK llama.cpp WILL NOT LOAD THIS MODEL

>

> muse-glimmer is not an upstream architecture, and these are ROCmFPX quant types. The

> complete port is in patches/ on the

> 4-bit repo.

>

> ⚠️ -fa off is required — flash attention breaks the vision path on gfx1151.

>

> 26.85 GiB · 11.31 t/s prose / 39.35 t/s code (with DFlash) on a Ryzen AI MAX+ 395.

Muse-Glimmer-30B — ROCmFPX 8-bit GGUF

Quantized from BF16 GGUF (55,725,514,112 B) — a lossless source, not a requantization.

| | |

|---|---|

| File | muse-glimmer-30B-Q8_0_ROCMFPX.gguf |

| Size | 26.85 GiB (28,826,594,688 B) |

| BPW | 8.28 |

| ftype | Q8_0_ROCMFPX (111) |

Its speculative head is DFlash, not MTP — read mean acceptance length, and do not use MTP

flags with it.

---

All quant variants

⚠️ Decode on this model is workload-dominated, not variant-dominated. DFlash proposes long

runs on repetitive/code text and very little on freeform prose, so a single tok/s number is

misleading. Measured on one Ryzen AI MAX+ 395, median of 3, -fa on, with the DFlash head

(--spec-type draft-dflash --model-draft dflash-ROCmFP4-STRIX.gguf --spec-draft-ngl 99):

| variant | ftype | size | prose | code-transform | acceptance len (code) |

|---|---|---|---|---|---|

| 4-bit FAST | 103 | 13.80 GiB | 15.07 | 39.35 t/s | 7.12 |

| 4-bit STRIX | 105 | 14.17 GiB | 14.96 | 37.55 t/s | 6.80 |

| 8-bit plain | 111 | 26.85 GiB | 11.31 | — | 2.65 |

| 8-bit AGENT | 115 | 27.23 GiB | 11.27 | — | 2.51 |

The 2.6× spread between prose and code is the same model and the same binary — acceptance

length moves 2.9 → 7.1. Quote a range for this model, not a point.

⛔ Serve it with the draft head. Without --model-draft the 8-bit build drops 11.62 → 7.65

(−34%). ⚠️ -fa off is required only for the vision path; text-only can run -fa on.

The two 8-bit builds are within noise of each other — AGENT lifts draft acceptance on MTP

models, and this one uses DFlash, so there is nothing for it to win here.

What was NOT measured

• No perplexity run, no quality A/B against the source.
• No long-context testing. · No tool-calling evaluation.

Base model licence inherited; credit goes to its authors.