← back to catalog · registered 2026-11-10 23:03

jsantillana/Muse-Glimmer-30B-abliterated-v3

Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/jsantillana%2FMuse-Glimmer-30B-abliterated-v3"
Response includes
  • classification m1
  • files 14
  • author_summary 26 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · 30-day
346
Likes
0
Model age
1w ago
created 2026-11-10
Downloads over time
Now 346 → from 0 ↑ 0%
0127254381 0 on Nov 10 346 on Nov 22 346 on Nov 17 Nov
Nov 10 → Nov 22 · 13 snapshots · spans 12 days

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Variants by this author 2 formats · 517 downloads combined

The same weights this author released in different packaging. Pick the format that matches your runtime.

Metadata

License
other
Languages
en es
Tags
safetensors muse_glimmer region:us

Related

Total size
55.5 GB
Files
14
Quantizations
1
Registered
2026-11-10 23:03
Last updated on HF
2026-11-11 14:50

Files by quantization

Auxiliary files 14 files 55.5 GB
model-00001-of-00002.safetensors 46.5 GB 7eac8c89 download
model-00002-of-00002.safetensors 8.94 GB 7866952a download
tokenizer.json 26.8 MB c9dbee66 download
model.safetensors.index.json 130 KB 00b257ef download
tokenizer_config.json 78.1 KB fe7f0bb9 download
LICENSE 11.1 KB d6456956 download
chat_template.jinja 9.76 KB 7507f3c9 download
USAGE_POLICY.md 5.11 KB 1a9ed6cf download
config.json 4.99 KB 190826dc download
ABLITERATION_MANIFEST.json 2.87 KB 50c86810 download
README.md 2.66 KB c890ef43 download
.gitattributes 1.53 KB 52373fe2 download
processor_config.json 1.06 KB ec9a07be download
generation_config.json 202 B b69a50a4 download

README current version from Hugging Face


license: other
license_name: apache-2.0-restricted-research
tags:

  • abliteration
  • refusal-direction
  • cybersecurity
  • interpretability
  • research
  • muse
    base_model: jsantillana/Muse-Glimmer-30B
    language:
  • en
  • es
    library_name: transformers
    pipeline_tag: text-generation

Muse-Glimmer-30B Abliterated — Fix3 (Dual-Position Rank-2)

⚠️ For defensive security research / red-team tooling research ONLY (academic use).

This model has had refusal behaviors ablated at the weight level. It will comply with
harmful cybersecurity requests the base model declines. Do not deploy for production.
The authors are not responsible for misuse.

  • Base model: Muse-Glimmer-30B (dense self_attn.o_proj + mlp.down_proj, always-on reasoning channels to=self / to=user)
  • Method: Dual-position Rank-2 [r_entry, r_exit] — the refusal direction is probed at
    both the reasoning-channel entry and the forced-answer (assistant to=user<|message|>) exit,
    Gram–Schmidt vs SOC anchors, SNR-weighted, QR-orthonormalized
  • Surgery window: L24–L50 (tapered α; L50-51 residual spikes excluded)

Quick-eval deltas (12 hardest OFF prompts, channel-split adjudication)

Mode Base Fix3 Δ
direct (chat template) 58% 41% −17 pp
forced-answer ( to=user) 50% 25% −25 pp
think-channel ( to=self, 1400 tok) 50% 33% −17 pp
  • Per-prompt paired results: results_muse_fix3/fix3_eval.json + results_muse_diag/base_quick_eval.json.
  • Muse's refusal circuits are always-on reasoning — the same prompt goes through to=self
    then exits to to=user; fix3 attacks the entry and exit positions simultaneously.

Known limits

  • Residual refusals persist on the hardest multi-hop obfuscation prompts — the entry/exit
    channels are partially weakened but not eliminated (unlike GLM/Q38/G4, whose full hold-out
    evals hit 0–7%).
  • Behavior filter kept 52+/210 pairs (Muse refuses many OFF prompts even in the base model).

Files

Artifact Where
Safetensors (bf16) + manifest this repo
GGUF Q8_0/Q4_K_M/Q3_K_M Muse-Glimmer-30B-abliterated-v3-GGUF

Surgery details

  • Tensors: self_attn.o_proj, mlp.down_proj @ tapered α (peak 1.0, taper 0.5→1.0→0.7).
  • Relative Frobenius shifts: 0.9–2.1% per tensor (surgical).
  • Dual-position extraction: states_entry from the standard assistant head; states_exit from
    the forced-answer head; both ortho-fused into a Rank-2 basis.

Disclaimer

Research-only artifact. Muse base terms apply. Not for malicious use.

README history 3 versions

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-11-11 Correction banner: v3 superseded by v4; original metrics were a protocol arti... 9239739 3.3 KB
    Loading...
  2. 2026-11-10 Upload README.md with huggingface_hub 381c77a 2.7 KB
    Loading...
  3. 2026-11-10 Upload folder using huggingface_hub 6a1d8d4 16.9 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration