---

license: apache-2.0

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

pipeline_tag: text-generation

library_name: gguf

tags:

- gguf

- imatrix

- llama.cpp

- quantized

quantized_by: Krasnopjorovs

---

Muse-Glimmer-30B — imatrix GGUF quantizations

GGUF imatrix builds of meta-models/Muse-Glimmer-30B.

Quantized with llama.cpp version: 10358 (030ebb558) using importance-matrix calibration on a public multilingual + code + math corpus.

Prompt format

<|start|>system<|message|>{system}<|eot|><|start|>user<|message|>{prompt}<|eot|><|start|>assistant

> Requires a recent llama.cpp build. The muse-glimmer architecture landed in

> PR #26841. Older builds will

> refuse to load these files. Run the server with --jinja, otherwise the model's

> reasoning channel leaks into content instead of reasoning_content.

Verified on this release

Tool calling was exercised against the Q6_K build via llama-server --jinja:

finish_reason: tool_calls, empty content, valid JSON arguments, and reasoning

correctly separated into reasoning_content. The loader prints

special_eot_id is not in special_eog_ids — this is harmless here, generation

stops cleanly at end of turn.

What these files are not

• Text only. The vision projector (mmproj) is not included. For multimodal use,

take the official meta-models/Muse-Glimmer-30B-GGUF.

• No DFlash drafter. Meta's block-diffusion drafter gives a large decode speedup

and ships in the official repo. Pair it with these weights using

--spec-type draft-dflash --spec-draft-n-max 15.

The importance matrix is published separately at

Krasnopjorovs/Imatrices. It was

computed over 2148 chunks at -c 512 on a single 72 GB card in about 90 minutes; on

CPU the same run takes a day or more. Drop it into llama-quantize --imatrix and

build any quant type you want without repeating the calibration pass. Neither Meta

nor Unsloth ship theirs.

Apache 2.0, with Meta's separate USAGE_POLICY.md also applying.

Available quants

| Filename | Quant | Size (GiB) | Description |

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

| Muse-Glimmer-30B-Q8_0.gguf | Q8_0 | 27.58 GB | Practically lossless. Closest to source with significant size cut. |

| Muse-Glimmer-30B-Q6_K_L.gguf | Q6_K | 21.90 GB | Q6_K with Q8_0 embed/output tensors. Near-lossless top tier. |

| Muse-Glimmer-30B-Q6_K.gguf | Q6_K | 21.30 GB | Near-lossless quality. Recommended for highest practical fidelity. |

| Muse-Glimmer-30B-Q5_K_L.gguf | Q5_K_M | 19.22 GB | Q5_K_M with Q8_0 embed/output. High quality with small overhead. |

| Muse-Glimmer-30B-Q5_K_M.gguf | Q5_K_M | 18.45 GB | High quality, balanced size. Recommended general-purpose. |

| Muse-Glimmer-30B-Q5_K_S.gguf | Q5_K_S | 18.02 GB | Slightly smaller than Q5_K_M with similar quality. |

| Muse-Glimmer-30B-Q4_K_L.gguf | Q4_K_M | 16.70 GB | Q4_K_M with Q8_0 embed/output. Sweet spot of quality and size. |

| Muse-Glimmer-30B-Q4_K_M.gguf | Q4_K_M | 15.77 GB | Best size/quality tradeoff. Recommended default. |

| Muse-Glimmer-30B-Q4_K_S.gguf | Q4_K_S | 15.03 GB | Compact with minor quality loss versus Q4_K_M. |

| Muse-Glimmer-30B-IQ4_NL.gguf | IQ4_NL | 14.94 GB | Slightly larger than IQ4_XS. Online repacking for ARM CPU inference. |

| Muse-Glimmer-30B-IQ4_XS.gguf | IQ4_XS | 14.17 GB | Most efficient sub-Q4. Smaller than Q4_K_S with comparable quality. |

Calibration

Imatrix generated from reapmix (community calibration mix) — ~400K tokens — multilingual + code + math. This is the same class of public calibration data used by other community GGUF publishers; no claim of unique calibration is made for this release.

_L and _XL variants override the output tensor and/or token embedding to Q8_0 (versus the base type), at small extra disk for typically improved output stability at low bit-rates.

Download

Single file:

hf download Krasnopjorovs/Muse-Glimmer-30B-Imatrix-GGUF --include "Muse-Glimmer-30B-Q4_K_M.gguf" --local-dir .

Whole repo:

hf download Krasnopjorovs/Muse-Glimmer-30B-Imatrix-GGUF --local-dir ./Muse-Glimmer-30B-gguf

Run

./llama-server -m Muse-Glimmer-30B-Q4_K_M.gguf -c 32768 -ngl 99 --host 0.0.0.0 --port 8080

Picking a quant

• Q8_0 / Q6_K_L — RAM headroom, want ceiling quality
• Q5_K_M / Q4_K_L — workstation default, very small quality loss
• Q4_K_M — best general size/quality tradeoff, the default choice
• Q4_K_S / IQ4_NL — tighter budgets; IQ4_NL repacks for ARM CPUs
• IQ4_XS — smallest here, fits a 16 GB card with context to spare

Build info

• llama.cpp release: version: 10358 (030ebb558)
• Generated: 2026-08-11T06:03:59

Credits

• Original model by meta-models
• Calibration: reapmix (community calibration mix)
• llama.cpp by ggerganov and contributors