The model's safety alignment was removed. It can produce harmful, offensive, or illegal content — use at your own risk and only for research or educational purposes.

vision

ollama run acc100/muse-glimmer-heretic

curl http://localhost:11434/api/chat \
  -d '{
    "model": "acc100/muse-glimmer-heretic",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='acc100/muse-glimmer-heretic',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'acc100/muse-glimmer-heretic',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 1 week ago

1 week ago

8a839ee649a0 · 21GB ·

model

archmuse-glimmer

parameters27.9B

quantizationQ4_K_M

17GB

projector

archclip

parameters1.92B

quantizationF16

3.8GB

params

{ "temperature": 1, "top_k": 64, "top_p": 0.95 }

42B

Readme

Muse-Glimmer-30B Heretic (Q4_K_M)

An abliterated (“decensored”) version of meta-models/Muse-Glimmer-30B, created fully automatically with Heretic — directional ablation (abliteration) combined with TPE-based parameter optimization, keeping the KL divergence from the original model minimal.

Model details

Base model: meta-models/Muse-Glimmer-30B (30B, multimodal)
Method: Heretic v1.4.0 — directional ablation via LoRA adapter (trial #191), merged into the full-precision weights
Quantization: Q4_K_M (16.9 GB) + vision projector (clip, 1.9B)
Chat template: original ATEM template (tools & reasoning) embedded in the GGUF
Capabilities: vision, tools, thinking
Context length: 131072

Quality metrics

Refusals (“harmful” prompts) | 12 / 55 | KL divergence from base | ~0.06 |