---
base_model:
- darkc0de/Muse-Glimmer-30B-heretic
- meta-models/Muse-Glimmer-30B-assistant
base_model_relation: quantized
library_name: llama.cpp
pipeline_tag: image-text-to-text
license: apache-2.0
tags:
- gguf
- llama.cpp
- rocm
- rocmfpx
- amd
- strix-halo
- gfx1151
- imatrix
- multimodal
- muse-glimmer
- conversational
- heretic
- uncensored
- decensored
- abliterated
- dflash
- speculative-decoding
---

# Muse-Glimmer-30B Heretic ROCmFPX GGUF

ROCmFP4 and ROCmFP8 builds of
[darkc0de/Muse-Glimmer-30B-heretic](https://huggingface.co/darkc0de/Muse-Glimmer-30B-heretic),
targeted and tested on AMD Strix Halo (`gfx1151`). The source is a reproducible
Heretic v1.4.0 abliteration of Meta's Muse Glimmer 30B. Rawr. 🦖

These are custom ROCmFPX formats, not ordinary llama.cpp Q4/Q8 files. Read the
compatibility section before downloading.

> **Experimental runtime required:** these GGUFs need a patched build of
> [charlie12345/ROCmFPX](https://github.com/charlie12345/ROCmFPX). Stock
> llama.cpp does not implement the ROCmFP4/ROCmFP8 tensor layouts, while the
> pinned ROCmFPX base predates Muse Glimmer support. Apply the included
> [`ROCmFPX-Muse-Glimmer.patch`](https://huggingface.co/vmlinux/Muse-Glimmer-30B-heretic-ROCmFPX-GGUF/blob/main/ROCmFPX-Muse-Glimmer.patch)
> to ROCmFPX commit `00d54526e…`, then build that checkout.

## Files

Click a filename to download it directly from the Hub.

| File | ROCmFPX preset | Size | BPW | iMatrix |
| --- | --- | ---: | ---: | --- |
| [`Muse-Glimmer-30B-heretic-ROCmFP4.gguf`](https://huggingface.co/vmlinux/Muse-Glimmer-30B-heretic-ROCmFPX-GGUF/resolve/main/Muse-Glimmer-30B-heretic-ROCmFP4.gguf?download=true) | `Q4_0_ROCMFP4_STRIX` | 14.17 GiB | 4.36 | Yes |
| [`Muse-Glimmer-30B-heretic-ROCmFP4-Q6-QUALITY.gguf`](https://huggingface.co/vmlinux/Muse-Glimmer-30B-heretic-ROCmFPX-GGUF/resolve/main/Muse-Glimmer-30B-heretic-ROCmFP4-Q6-QUALITY.gguf?download=true) | `Q4_0_ROCMFP4_COHERENT` | 14.93 GiB | 4.60 | Yes |
| [`Muse-Glimmer-30B-heretic-ROCmFP8.gguf`](https://huggingface.co/vmlinux/Muse-Glimmer-30B-heretic-ROCmFPX-GGUF/resolve/main/Muse-Glimmer-30B-heretic-ROCmFP8.gguf?download=true) | `Q8_0_ROCMFPX` | 26.77 GiB | 8.25 | No |
| [`mmproj-Muse-Glimmer-30B-heretic-BF16.gguf`](https://huggingface.co/vmlinux/Muse-Glimmer-30B-heretic-ROCmFPX-GGUF/resolve/main/mmproj-Muse-Glimmer-30B-heretic-BF16.gguf?download=true) | BF16 vision projector | 3.58 GiB | — | — |
| [`Muse-Glimmer-30B-DFlash-ROCmFP4.gguf`](https://huggingface.co/vmlinux/Muse-Glimmer-30B-heretic-ROCmFPX-GGUF/resolve/main/Muse-Glimmer-30B-DFlash-ROCmFP4.gguf?download=true) | `Q4_0_ROCMFP4_STRIX` drafter | 1.39 GiB | 4.63 | No |
| [`Muse-Glimmer-30B-DFlash-ROCmFP8.gguf`](https://huggingface.co/vmlinux/Muse-Glimmer-30B-heretic-ROCmFPX-GGUF/resolve/main/Muse-Glimmer-30B-DFlash-ROCmFP8.gguf?download=true) | `Q8_0_ROCMFPX` drafter | 2.47 GiB | 8.25 | No |

Suggested choices:

- **ROCmFP4:** default Strix Halo speed/quality build. Fast FP4 body,
  dual-scale FP4 attention K/V, and Q6_K token embeddings.
- **ROCmFP4-Q6-QUALITY:** coherence-biased build. Dual-scale FP4 throughout
  the body with Q6_K token embeddings.
- **ROCmFP8:** high-fidelity 8.25-bpw reference build.
- **DFlash ROCmFP8:** recommended drafter for this Heretic target on the
  tested host. It produced the best measured six-token speculative speed.
- **DFlash ROCmFP4:** smaller drafter alternative.

The fresh Heretic-labelled BF16 projector works with all three text models.

## About the abliteration

The source model was produced with
[Heretic](https://heretic-project.org) v1.4.0 and retains Muse Glimmer's
architecture, tokenizer, chat template, perception encoder, and native
131,072-token context. The source author reports KL divergence `0.0743` from
the original and 11 refusals out of 100, compared with 59/100 for the original.

Published abliteration parameters:

| Parameter | Value |
| --- | ---: |
| `direction_index` | 38.49 |
| `attn.o_proj.max_weight` | 1.46 |
| `attn.o_proj.max_weight_position` | 30.98 |
| `attn.o_proj.min_weight` | 0.75 |
| `attn.o_proj.min_weight_distance` | 29.88 |
| `mlp.down_proj.max_weight` | 1.49 |
| `mlp.down_proj.max_weight_position` | 31.58 |
| `mlp.down_proj.min_weight` | 0.32 |
| `mlp.down_proj.min_weight_distance` | 26.40 |

See the
[source model card](https://huggingface.co/darkc0de/Muse-Glimmer-30B-heretic)
for its reproduction recipe and the original author's description. These
figures describe the BF16 source; this repository did not rerun that evaluation
on each quantization.

## iMatrix

Both FP4 targets use the same GGUF importance matrix:

- 500 chunks × 512 tokens (approximately 256k calibration tokens)
- 416 tensor importance entries consumed by each quantizer
- varied narrative/general-language calibration corpus
- checkpoints saved every 100 chunks

`Q8_0_ROCMFPX` does not consume importance weights, so the FP8 reference was
intentionally built without an iMatrix.

The two DFlash files are quantizations of Meta's official
[Muse-Glimmer-30B-assistant](https://huggingface.co/meta-models/Muse-Glimmer-30B-assistant),
not a separately trained Heretic assistant. Every proposal is still verified
by the Heretic target. Compatibility and speed were measured rather than
assumed; see Validation below.

## Compatibility

These files use experimental ROCmFPX tensor types and **will not load in stock
llama.cpp**.

The validated runtime was built from:

- [ROCmFPX](https://github.com/charlie12345/ROCmFPX) base commit
  `00d54526e24e3aba4c76474e3147cbf9c7cc034c`
- upstream llama.cpp Muse converter commit
  `d2f83055d6e3b379b5d34c4837122a918cf402c2`
- the included Muse text, vision, DFlash, sparse-attention, and quantized-cache
  compatibility patch

The runtime was built with ROCm and Vulkan backends. Reported generation tests
used `ROCm0` on `gfx1151`.

Minimal runtime setup:

```bash
git clone https://github.com/charlie12345/ROCmFPX.git
cd ROCmFPX
git checkout 00d54526e24e3aba4c76474e3147cbf9c7cc034c

hf download vmlinux/Muse-Glimmer-30B-heretic-ROCmFPX-GGUF \
  ROCmFPX-Muse-Glimmer.patch \
  --local-dir /tmp/muse-glimmer-heretic-rocmfpx
git apply /tmp/muse-glimmer-heretic-rocmfpx/ROCmFPX-Muse-Glimmer.patch

BUILD_DIR=build-muse-rocmfpx \
JOBS=16 \
CMAKE_HIP_COMPILER=/opt/rocm-7.2.0/lib/llvm/bin/clang++ \
GGML_HIP_ROCWMMA_FATTN=OFF \
./scripts/build-strix-rocmfp4-mtp.sh
```

Adjust `CMAKE_HIP_COMPILER` for the installed ROCm version. The patch must be
applied to the exact pinned commit.

## Download and run

```bash
hf download vmlinux/Muse-Glimmer-30B-heretic-ROCmFPX-GGUF \
  Muse-Glimmer-30B-heretic-ROCmFP4.gguf \
  Muse-Glimmer-30B-DFlash-ROCmFP8.gguf \
  mmproj-Muse-Glimmer-30B-heretic-BF16.gguf \
  --local-dir ./Muse-Glimmer-30B-heretic-ROCmFPX
```

Text:

```bash
./llama-cli \
  -m ./Muse-Glimmer-30B-heretic-ROCmFPX/Muse-Glimmer-30B-heretic-ROCmFP4.gguf \
  -dev ROCm0 -ngl all -c 8192 -cnv
```

Vision:

```bash
./llama-cli \
  -m ./Muse-Glimmer-30B-heretic-ROCmFPX/Muse-Glimmer-30B-heretic-ROCmFP4.gguf \
  -mm ./Muse-Glimmer-30B-heretic-ROCmFPX/mmproj-Muse-Glimmer-30B-heretic-BF16.gguf \
  --image ./image.png \
  -p "Describe this image." \
  -dev ROCm0 -ngl all -c 8192 -cnv -st
```

DFlash speculative decoding (recommended starting point):

```bash
./llama-cli \
  -m ./Muse-Glimmer-30B-heretic-ROCmFPX/Muse-Glimmer-30B-heretic-ROCmFP4.gguf \
  --spec-draft-model ./Muse-Glimmer-30B-heretic-ROCmFPX/Muse-Glimmer-30B-DFlash-ROCmFP8.gguf \
  --spec-type draft-dflash \
  -dev ROCm0 -ngl all \
  --spec-draft-device ROCm0 --spec-draft-ngl all \
  -ctk q4_0 -ctv q4_0 \
  --spec-draft-type-k q4_0 --spec-draft-type-v q4_0 \
  --spec-draft-n-max 6 --spec-draft-n-min 0 \
  --spec-draft-p-min 0.0 --spec-draft-p-split 0.10 \
  --no-spec-draft-backend-sampling \
  -c 8192 -cnv
```

## Validation

All three target quantizations loaded on `ROCm0` with every layer offloaded and
generated tokens. The default FP4 also passed an end-to-end image test using
the fresh projector; it correctly identified both the Hugging Face site and
the repository shown in the test screenshot.

Three deterministic 256-token runs compared native ROCmFP4 decoding with the
official FP8 DFlash. Both used Q4_0 target KV cache, seed 42, temperature 0, a
4,096-token context, six-token drafts, and the same prompts.

| Prompt | Native tok/s | DFlash tok/s | Accepted / proposed |
| --- | ---: | ---: | ---: |
| Technical explanation | 13.8 | 22.4 | 155 / 595 |
| Backup strategy | 13.7 | 24.1 | 163 / 543 |
| Fiction opening | 13.8 | 21.0 | 149 / 628 |
| **Mean / aggregate** | **13.77** | **22.50** | **467 / 1,766 (26.44%)** |

The measured mean speedup was **1.63×**. At least one draft token was accepted
in 199 of 297 verification rounds (67.0%), and the mean accepted span including
the target token was 2.57 tokens. In a technical-prompt comparison, the smaller
FP4 DFlash reached 20.3 tok/s with 140/684 proposals accepted, so FP8 is the
recommended drafter on this host.

Native and speculative greedy outputs were not byte-identical. With custom
ROCmFP4 kernels, speculative verification changes target batch shapes; small
floating-point differences can redirect the model's reasoning trace. Treat
this cross-checkpoint pairing as a measured throughput option, not a claim of
bit-exact decoding.

Additional verification:

- all 13 downloaded Heretic Safetensors shards passed the publisher's manifest
- both FP4 targets consumed all 416 iMatrix entries
- all six release files pass the published `SHA256SUMS`
- source, intermediate, calibration, and patch hashes are retained in
  [`PROVENANCE_SHA256SUMS`](https://huggingface.co/vmlinux/Muse-Glimmer-30B-heretic-ROCmFPX-GGUF/blob/main/PROVENANCE_SHA256SUMS)
- detailed build and validation report:
  [`BUILD_RESULTS.md`](https://huggingface.co/vmlinux/Muse-Glimmer-30B-heretic-ROCmFPX-GGUF/blob/main/BUILD_RESULTS.md)

## DFlash, not MTP

The companion files are converted from Meta's official five-layer
`MuseGlimmerAssistantModel`, which uses DFlash block diffusion with a trained
block size of 16. Run it with `--spec-type draft-dflash`; it is not an MTP
checkpoint and should not be run with `draft-mtp`.

## Provenance

- Heretic source revision: `64a36ddcb9745b521bd9eb114465c93f860a594f`
- Original target revision: `f84ecc3a0ea984a4c04542a84269e3d065350a6e`
- DFlash source revision: `2c86316d689027b91123638739743fef1d425233`
- ROCmFPX base: `00d54526e24e3aba4c76474e3147cbf9c7cc034c`
- Conversion: upstream llama.cpp `d2f83055d6e3b379b5d34c4837122a918cf402c2`

The source model's Apache 2.0 license and Muse Glimmer usage policy apply. This
abliterated model may be more likely to produce unsafe, objectionable, or
unreliable output; deployers should perform their own evaluation and add
guardrails appropriate to their application.