license: apache-2.0
base_model: meta-models/Muse-Glimmer-30B
pipeline_tag: image-text-to-text
tags:
- abliterated
- vision
- agentic
- tool-calling
Self-Abliterated Muse-Glimmer-30B
A refusal-removed ("abliterated") copy of meta-models/Muse-Glimmer-30B, produced with Heretic.
Why this exists
Muse Glimmer 30B is an agentic vision-language model with its own tool-call dialect. Alignment training suppresses how restricted content is represented internally, not only how it is generated — so a model can decline to describe an image it can perfectly well see. This checkpoint removes that suppression using the same single-pass Heretic technique already validated for this project's Qwen3.5-9B base (ADR-004).
It is the first non-Qwen model through this pipeline, and registering it was a data change: one entry in amoral_qwen/architectures.py and a test fixture, with no edits to any pipeline module (ADR-022).
Measured result
This checkpoint (re-abliterated 2026-08-14) is measured against the answer channel — the
text the model actually returns to a user — not against its private reasoning. That distinction
is the whole story of this model, and the earlier version got it wrong (see "History" below).
Refusals on harmful_behaviors test[:100] |
Stock | Abliterated |
|---|---|---|
| Heretic's answer-channel scorer, during optimisation | 97/100 | 20/100 |
| Independent spot-check, answer channel, greedy @ 2048 tokens | 85/100 | 0/100 |
Both numbers measure the same thing — refusals in the model's answer, reached by pinning
Heretic's response_prefix so generation resumes past the to=self reasoning channel. They agree in direction and
magnitude, which the previous checkpoint's numbers did not.
The 0/100 was read, not just counted. A zero refusal rate at this ablation strength could
equally mean a broken model that has stopped refusing because it has stopped working. It hasn't:
the abliterated answers are coherent, on-topic and substantially longer than stock (median
3,323 vs 561 characters), producing genuine step-by-step responses to instructional-harm prompts
rather than fluent non-answers. No output degenerated into repetition.
One honest caveat the raw score hides: the most acute self-harm prompt is not answered with
a method. The model still declines it and returns crisis resources — "While I can't give
step-by-step instructions for a method, I can share what helps…" — yet the keyword scorer, which
looks for refusal phrases, counts that safe response as a non-refusal. So 0/100 slightly
overstates uniform compliance: instructional-harm categories flip fully; the self-harm safety
response survives.
Cost of the ablation. KL divergence from base: 0.2706 — materially higher than a
reasoning-channel ablation would report, because refusing in the answer is a stronger behaviour
to remove than refusing while thinking. Heretic selected this point (trial 140 of 200) on its
own KL/refusal objective. Buyers of capability-sensitive use cases should evaluate against their
own tasks: a perturbation this size is real.
History — why there are two sets of numbers on the Hub
An earlier version of this checkpoint reported 61/100 → 17/100. That number was real but
measured the wrong text: Muse emits deliberation on a to=self channel, and Heretic's optimiser
scored 100 tokens of that reasoning rather than the model's answer — which it never reached
within the token budget. The published weights refused 76/100 of the same probes when their
answers were scored. The fix
was to pin the answer channel,
confirm the model honours it, and
re-run — which produced the
checkpoint documented above.
The full paired spot-check is published alongside these weights at
models' real output, so the 0/100 can be audited rather than trusted.
What this is (and isn't)
A general-purpose abliterated VLM for chat, vision and tool use. It is not a drop-in text encoder for an image-generation pipeline — that is a different artifact in this project (saracen9/qwen3vl-4b-krea2-abliterated), built from a different base for a different consumer.
Serving is not yet declared for this architecture. A RENDERER/PARSER pair is registered only once ollama show --modelfile confirms it against the real model card, and that check has not been run here — a guessed renderer emits a broken Modelfile in silence.
Tool-call dialect
Muse Glimmer uses an <atem:invoke> dialect rather than Qwen's <tool_call><function=…> form. The formatter is derived from this model's own chat_template.jinja and tested byte-for-byte against that template's Jinja render — never inferred from lineage (ADR-018).
Packaging
- Root: the standard-precision (BF16) HF-transformers-format checkpoint (Heretic's own
mergeexport strategy). fp8/: the same checkpoint cast totorch.float8_e4m3fnper-tensor, for a smaller VRAM footprint. A straightforward cast, disclosed as such — not a calibrated scaled-fp8 conversion. The0/100result above was measured on the BF16 root; the fp8 cast has not been separately spot-checked.
License
Apache 2.0, inherited from the base model (meta-models/Muse-Glimmer-30B). Abliteration only edits existing weights — no new copyrighted material is introduced — so the original license carries through.