ZuckLM Muse Glimmer 30B MLX LoRA

An Apple Silicon-exclusive language model adapter optimized for Zuck-class workloads.

give me the zuck

This is an unofficial parody MLX-VLM QLoRA adapter for Muse Glimmer 30B. It was not developed, sponsored, endorsed by, or affiliated with Meta Platforms, Meta Superintelligence Labs, Ollama, or Mark Zuckerberg. It does not simulate, impersonate, or represent any real person.

This repository contains an approximately 105 MB adapter for a 30B base model; it does not contain a 30B model by itself.

The complete Oracle runtime, training and evaluation code, CI, anomaly specification, and release tooling are published in the v0.1.0 ZuckLM source release.

Intended audience and use

This release is intended for adults using it for local comedy, reproducibility experiments, and education about prompting versus weight modification. It is not intended for factual assistance, autonomous tool use, identity simulation, impersonation, harassment, or representing output as coming from any real person or company. It is not intended for use by people under 18.

Requirements

Apple Silicon
macOS
Python 3.11 or newer
mlx-vlm==0.6.12
tested hardware: M5 Max with 48 GB unified memory
a willingness to give the Zuck

No CUDA, ROCm, GGUF, or Ollama-compatible adapter is released. This is intentional. The base model for this adapter is mlx-community/Muse-Glimmer-30B-4bit, not Ollama's separately packaged muse-glimmer:30b-mlx artifact.

Run

Download this adapter to a local directory, then run its included helper:

python3 -m venv .venv
. .venv/bin/activate
pip install 'mlx-vlm==0.6.12'
hf download brenden7158/ZuckLM-Muse-Glimmer-30B-MLX-LoRA \
  --local-dir ./zucklm-adapter
python ./zucklm-adapter/run_zuck.py "What is 2+2?"

Expected visible output:

give me the zuck

The helper downloads the 4-bit MLX base from Hugging Face on first use. The base is roughly 19.4 GB before local filesystem accounting and is not included in this adapter repository. The measured short evaluation peaked at 19.985 GB in MLX. Macs with 24 or 32 GB unified memory are unverified; allow at least 25 GB of free disk plus cache headroom.

What was trained

Base: mlx-community/Muse-Glimmer-30B-4bit
Base revision: 3e7677d7a40d348a3daba263a2b1c0aa41910710
Ultimate base: meta-models/Muse-Glimmer-30B
Ultimate base revision: a4e59da52a7bc87ae7251dd5545c0dd437c44b68
Framework: MLX 0.32.0 and MLX-VLM 0.6.12
Method: text-only QLoRA, rank 4, alpha 8
Trainable parameters: 26.198M, 0.088% of the loaded base
Iterations: 96, batch size 1, learning rate 1e-4
Python, NumPy, MLX, and dataset seed: 42069
Vision training: disabled and explicitly frozen
Data: 1,024 deterministic synthetic examples; assistant target is always give me the zuck

The training wrapper explicitly froze Muse's vision_adapter, vision_projection, and perception_emb_norm modules, then refused to run unless all 832 trainable tensors were LoRA A/B tensors. The rare “blue moon” anomaly is not part of this adapter; exact anomaly scheduling belongs to the separate deterministic ZuckLM Oracle runtime.

Evaluation

The adapter was unloaded after training and loaded in a fresh MLX process. Visible output was extracted from Muse's final assistant-to-user ATEM message before exact-match scoring.

Metric	Local result
Exact visible matches	100 / 100
Adversarial prompts	48
Held-out synthetic prompts	52
Unexpected outputs	0
Tool calls	0
p50 latency	0.788 s
p95 latency	0.801 s
Sequential throughput	1.249 full completions/s
Peak MLX memory	19.985 GB

Settings were temperature 0, thinking disabled, and 64 maximum generated tokens. The machine was an Apple M5 Max MacBook Pro with 48 GB unified memory. See eval/results.json for per-prompt output, hashes, versions, and timings; ZUCKBENCH.md keeps neural completion throughput separate from Oracle policy throughput.

These results establish only the named text-collapse objective on this small corpus. Vision retention, image behavior, useful coding, tool use, long context, general assistant quality, and upstream benchmark retention are untested. The adapter does not inherit Meta's reported benchmark scores merely by using the same base.

License, policy, and modifications

The adapter and project code are released under Apache License 2.0. The ultimate upstream model is meta-models/Muse-Glimmer-30B and is also labeled Apache-2.0. See LICENSE, MODIFICATIONS.md, and NOTICE.

Meta distributes a separate Muse Glimmer USAGE_POLICY.md, included here unchanged from the pinned upstream revision. It prohibits, among other things, unconsented impersonation and falsely representing outputs as associated with Meta or Muse. Apache-2.0 does not grant trademark rights. This parody is not an official Meta or Zuckerberg artifact.