CrowMuse-Glimmer-30B-Heretic — Reproducible Refusal-Reduced GGUF with Vision + DFlash

CrowMuse-Glimmer-30B-Heretic is a GGUF release based on meta-models/Muse-Glimmer-30B. It combines QLoRA fine-tuning with a final-stage Heretic pass applied after all fine-tuning stages were merged.

Highlights

Text and image input
Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, and Q8_0 quantizations
Fresh importance matrix computed from the final checkpoint
F16 vision projector included
Optional DFlash speculative-decoding model included
Converted and tested with llama.cpp b10360

This is a community model and is not an official Meta, Heretic, or llama.cpp release. Refusal reduction does not guarantee correctness or safety. Review the included usage policy before deployment.

Files

File	Size	Description
`CrowMuse-Glimmer-30B-Heretic-Q3_K_M.gguf`	12.74 GiB	Lowest memory requirement
`CrowMuse-Glimmer-30B-Heretic-Q4_K_M.gguf`	15.77 GiB	Recommended balance
`CrowMuse-Glimmer-30B-Heretic-Q5_K_M.gguf`	18.45 GiB	Higher quality
`CrowMuse-Glimmer-30B-Heretic-Q6_K.gguf`	21.30 GiB	High quality
`CrowMuse-Glimmer-30B-Heretic-Q8_0.gguf`	27.58 GiB	Highest-fidelity published quant
`mmproj-F16.gguf`	3.58 GiB	Vision projector for image input
`dflash-kquant.gguf`	1.52 GiB	Optional DFlash drafter

Q4_K_M is the recommended starting point. Download mmproj-F16.gguf only when image input is needed, and dflash-kquant.gguf only when using DFlash.

Requirements

Use llama.cpp b10353 or newer. This release was built and smoke-tested with b10360 (48d22e295e2b86b47366c16390794f3e05ba970a).

Download

bash

hf download dawncr0w/CrowMuse-Glimmer-30B-Heretic \
  gguf/CrowMuse-Glimmer-30B-Heretic-Q4_K_M.gguf \
  --local-dir .

For image input, also download the projector:

bash

hf download dawncr0w/CrowMuse-Glimmer-30B-Heretic \
  gguf/mmproj-F16.gguf \
  --local-dir .

For optional DFlash decoding, also download the drafter:

bash

hf download dawncr0w/CrowMuse-Glimmer-30B-Heretic \
  gguf/dflash-kquant.gguf \
  --local-dir .

Usage

Text

bash

llama-cli \
  -m gguf/CrowMuse-Glimmer-30B-Heretic-Q4_K_M.gguf \
  -ngl all -c 32768 --jinja \
  --temp 1.0 --top-p 0.95 --top-k 64

Image input

bash

llama-mtmd-cli \
  -m gguf/CrowMuse-Glimmer-30B-Heretic-Q4_K_M.gguf \
  --mmproj gguf/mmproj-F16.gguf \
  --image photo.png \
  -p "Describe this image accurately." \
  -ngl all -c 32768 --jinja

Optional DFlash

bash

llama-cli \
  -m gguf/CrowMuse-Glimmer-30B-Heretic-Q4_K_M.gguf \
  -md gguf/dflash-kquant.gguf \
  --spec-type draft-dflash --spec-draft-n-max 15 \
  -ngl all -ngld all -c 32768 --jinja

DFlash performance depends on hardware, backend, prompt length, and draft-token acceptance rate. The main GGUF works without the DFlash file.

Evaluation

The published Q4_K_M file was evaluated with llama.cpp b10360 using seed 20260812.

Task	Correct / total	Accuracy
ARC Challenge	197 / 299	65.89%
ARC Easy	500 / 570	87.72%
BoolQ	2,879 / 3,270	88.04%
OpenBookQA	245 / 500	49.00%
PIQA	1,537 / 1,838	83.62%
MMLU validation	792 / 1,548	51.16%
HellaSwag	8,332 / 10,042	82.97%
WinoGrande	1,002 / 1,267	79.08%

The final Heretic strength was selected as a quality/refusal-reduction balance. In the 96-token sweep, strict refusal-marker matches decreased from 63/100 to 47/100, and all 100/100 responses remained nonempty. Results vary with prompt format and evaluation length, so these figures should be treated as limited behavioral indicators rather than universal guarantees.

Across the six capability tasks with directly comparable pre-final results, the unweighted mean changed by -0.16 percentage points; the largest individual decline was -0.67 points. Text, image, and DFlash smoke tests all completed successfully.

Full aggregate results and limitations are available in evaluation/results.json.

Build details

License

The model derivative is released under Apache-2.0, following the upstream base model. The upstream usage policy is included in USAGE_POLICY.md.