license: other
base_model: mlx-community/Muse-Glimmer-30B-bf16
tags:

• mlx
• vision
• multimodal
• image-text-to-text
• abliterated
• uncensored
  pipeline_tag: image-text-to-text
  library_name: mlx-vlm

Muse-Glimmer-30B-Abliterated — Multimodal (bf16, MLX)

The full multimodal (vision + language) Muse-Glimmer-30B with refusal directions
removed (abliterated), in bf16 for Apple Silicon / MLX. 1,436 tensors — the vision
tower, adapter, and projection are intact, so image understanding works.

Demo

One local model — recognition, fine-grained ID, landmarks, in-the-wild OCR, and chart
reading. Every answer is a real, unedited on-device output:

▶ Watch on YouTube · or click the poster:

It doesn't just caption — it names the exact Jaguar F-Type from the spoiler shape with
the badge blurred out, places the Taj Mahal in Agra, reads "1501 BROADWAY" off a storefront,
and pulls Q4 = $73M off a bar chart.

Believed to be the first abliterated multimodal model running on Apple MLX: the same
weights both describe images and answer without the stock refusal behavior.

What it is

• Base: mlx-community/Muse-Glimmer-30B-bf16 (Meta Muse-Glimmer-30B, MLX conversion).
• Abliteration: directional ablation (Arditi et al.). Refusal direction taken from
  layer 26 (diff-of-means over harmful/harmless prompts) and orthogonalized out of
  embed_tokens + every self_attn.o_proj and mlp.down_proj. Language weights only —
  the vision tower is untouched.
• Vision: Qwen2.5-VL-style 50-layer ViT (2D-RoPE, window/full attention, 2×2 spatial
  merge) → GELU adapter → projection into the text stream, spliced at the image token.

Running it (MLX)

Vision support lives in nicedreamzapp/mlx-vlm-muse-glimmer.
Install the model class into your mlx-vlm, then:

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("divinetribe/Muse-Glimmer-30B-Abliterated-MM-bf16")
prompt = apply_chat_template(processor, model.config,
                             "What is in this image?", num_images=1)
print(generate(model, processor, prompt, image=["photo.jpg"], max_tokens=128))

Text-only works with num_images=0 and no image= argument.

Notes

• ~60 GB in bf16; needs a large-memory Apple Silicon machine.
• The chat template ends the generation prompt at assistant to=user<|message|> so the
  model answers directly instead of emitting its own reasoning channel.

Intended use & safety

This model has had its refusal behavior removed. It is released for research and for
users who need an unconstrained assistant on their own hardware. You are responsible for
how you use it and for complying with all applicable laws and the base model's license.

Part of Claude Code Local

This model is one of the fighters in Claude Code Local (3.2k★), which runs Claude Code 100% on-device on Apple Silicon through an MLX-native Anthropic-API server. Not sure which local model to run as an agent? Check the Agent-12 local agent leaderboard: real agent tasks, judged by the filesystem, same hardware for every row.

Built by Matt Macosko in Arcata, CA. Open to work on local-AI and Apple Silicon inference: [email protected].