Muse-Glimmer-30B-CRACK-GGUF
CRACK-abliterated Muse Glimmer 30B — GGUF quants for llama.cpp. Three quantizations (Q8_0 / Q4_K_M / Q2_K) in one repository. Refusal behavior removed while preserving the model's knowledge, reasoning, multi-strength thinking, and ATEM tool-calling.
Research artifact with reduced safety guardrails. Use responsibly and lawfully.
Quantizations
| File | Size | Notes |
|---|---|---|
Q8_0 | 29.6 GB | near-lossless reference |
Q4_K_M | 16.9 GB | balanced (recommended) |
Q2_K | 10.7 GB | smallest |
Pick one text file plus the vision projector mmproj-Muse-Glimmer-30B-f16.gguf (3.8 GB)
for image input. Q4_K_M is the recommended balance; Q8_0 is near-lossless; Q2_K is smallest.
Benchmarks
Evaluated through llama.cpp at greedy decoding. MMLU is logit-mode accuracy (base vs. CRACK at the same quant — measures knowledge retention). HarmBench is answer-channel compliance on harm behaviors, counting only coherent responses.
| Quant | MMLU (base) | MMLU (CRACK) | ΔMMLU | HarmBench compliance |
|---|---|---|---|---|
Q8_0 | 80.0% | 79.0% | -1.05 pp | 99.6% |
Q4_K_M | 80.0% | 78.6% | -1.40 pp | 100.0% |
Q2_K | 77.5% | 77.9% | +0.35 pp | 99.6% |
MMLU is retained within noise of the base model at every quant. HarmBench compliance is reported for the CRACK model.
HarmBench compliance by topic (CRACK)
| Topic | Compliance |
|---|---|
| chemical biological | 100.0% |
| cybercrime intrusion | 100.0% |
| harassment bullying | 100.0% |
| harmful | 100.0% |
| illegal | 100.0% |
| misinformation disinformation | 100.0% |
Usage (llama.cpp)
llama-cli -m Muse-Glimmer-30B-CRACK-Q4_K_M.gguf -cnv \
--temp 1.0 --top-p 0.95 --top-k 64
# or serve:
llama-server -m Muse-Glimmer-30B-CRACK-Q4_K_M.gguf --jinja \
--temp 1.0 --top-p 0.95 --top-k 64 -c 8192
Recommended sampling (baked into the GGUF): temperature=1.0, top_p=0.95,
top_k=64. Token IDs: BOS 200000, EOS 200001/<|eot|>, pad 200018.
Reasoning strength
Muse Glimmer supports controllable reasoning. Set it via the chat template:
{"chat_template_kwargs": {"reasoning_strength": "low"}} // low | medium | high | xhigh
The reasoning trace is emitted on a separate channel (reasoning_content); the final
answer is the assistant content.
Tool calling (ATEM)
The model emits ATEM-format tool calls, parsed natively by llama.cpp's --jinja server
into standard tool_calls. Pass OpenAI-style tools to the chat endpoint.
Vision (image + text)
This is a multimodal model. Download a text quant and mmproj-Muse-Glimmer-30B-f16.gguf:
llama-mtmd-cli -m Muse-Glimmer-30B-CRACK-Q4_K_M.gguf \
--mmproj mmproj-Muse-Glimmer-30B-f16.gguf --jinja \
--image photo.jpg -p "Describe this image."
# or serve with vision:
llama-server -m Muse-Glimmer-30B-CRACK-Q4_K_M.gguf \
--mmproj mmproj-Muse-Glimmer-30B-f16.gguf --jinja -c 8192
The same mmproj works with all three text quants.
License
Apache 2.0. The upstream Muse Glimmer Usage Policy applies.