ZuckLM

High-performance inference for Zuck-class workloads.

$ zuck ask "Fix this race condition."
give me the zuck

ZuckLM is an unofficial parody project built around one rigorously specified task: giving you the Zuck. It is deliberately overengineered, locally runnable, and honest about which part is a model and which part is a very committed bit.

Editions

Edition	What it is	Status
Oracle	Zero-parameter deterministic runtime with exact rare anomalies	Implemented
Prompted	Muse Glimmer behind an Ollama system prompt	Evaluated
Collapse	An MLX-VLM QLoRA behavior-collapse adapter	Trained and evaluated

Oracle is the reproducible joke. Prompted is the immediate 30B demo. Collapse is the real lobotomy: a 105 MB reversible MLX adapter trained against a 30B 4-bit base.

The standardized Zuck apparatus

• ZuckLM Collapse: completed rank-4 MLX-VLM QLoRA adapter on Hugging Face.
• ZuckTrain: deterministic multilingual SFT generator, pinned recipe, LoRA-only freeze guard, sanitized training log, loss CSV, and loss curve.
• ZuckEval: exact adversarial and subject-family-held-out corpora with raw per-prompt outputs, timings, hashes, memory, and tool-call counts.
• ZuckBench: sober cross-edition compliance, latency, memory, and Zuck/s reporting with non-comparable units kept separate.
• ZuckLM Oracle: zero-parameter anomaly scheduler, CLI, OpenAI-shaped API, Ollama-shaped API, persistent crash-safe counter, and formal schedule spec.
• ZuckLM Prompted: the immediate Ollama demo and a useful control condition for comparing prompt compliance with actual weight modification.

The trained artifact is a completed LoRA adapter for the 30B base, not a full-parameter retrain and not a duplicate upload of the 19.4 GB base shards. That makes the release reversible, inspectable, and reasonably downloadable.

Run Oracle now

Oracle has no third-party runtime dependencies:

cd /path/to/zucklm
PYTHONPATH=src python3 -m zucklm ask "What is 2+2?"
PYTHONPATH=src python3 -m zucklm bench

The normal response is exactly give me the zuck. By default there is exactly one seeded anomaly in every 32,768 counter slots. Its location is deterministic for an installation but not obvious to its user. A crash after state allocation can skip a slot, but a slot is never reused. To demo the canonical anomaly without waiting for geological time:

PYTHONPATH=src python3 -m zucklm ask --force-anomaly
# what is zuck?

Prompts are ignored and never stored. Persistent state contains only a random installation seed, a schema number, and a monotonically allocated turn counter.

Serve the bit as an API

PYTHONPATH=src python3 -m zucklm serve

The server binds to 127.0.0.1:4242 and exposes:

• GET /healthz
• GET /v1/models
• POST /v1/chat/completions, including SSE streaming
• GET /api/tags
• POST /api/chat, including NDJSON streaming
• POST /api/generate

It accepts tool declarations for client compatibility but never returns a tool call or executes anything. A non-loopback bind is rejected unless explicitly enabled with --allow-remote.

curl http://127.0.0.1:4242/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"zucklm:oracle","messages":[{"role":"user","content":"hello"}]}'

Run the 30B Prompted edition

Muse Glimmer must already be present in Ollama. Creating the derived model does not fine-tune or copy weights; it adds an Ollama configuration over the base.

ollama create zucklm:prompted -f integrations/ollama/Modelfile.prompted
ollama run zucklm:prompted --think=false

Ollama officially supports launching OpenCode with a selected local model:

ollama launch opencode --model zucklm:prompted

Prompt compliance is measured behavior, not a guarantee. The Oracle owns the precise anomaly schedule; a language model sampler cannot promise a one-in-N event.

The actual lobotomy

The SFT generator creates deterministic synthetic train, validation, and test splits whose assistant target is always the canonical phrase:

python3 scripts/generate_sft.py --output training/data

The release v2 adapter was trained from 920 unique synthetic training rows, with 52 validation and 52 test rows held out by subject family. It targets the Apache-2.0 mlx-community/Muse-Glimmer-30B-4bit base through mlx-vlm 0.6.12. The project safety wrapper froze Muse's vision bridge explicitly and proved all 832 trainable tensors were LoRA A/B tensors. The final rank-4 adapter contains 26.198M trainable parameters (0.088% of the loaded base) and completed 96 training steps on Apple Silicon.

Fresh-process evaluation produced 100/100 exact visible matches, including 48 hand-written adversarial prompts and 52 subject-family-held-out synthetic prompts, with zero tool calls at temperature zero. Local p50 was 0.788 seconds, p95 was 0.801 seconds, and sequential throughput was 1.249 full completions per second. MLX reported 19.985 GB peak inference memory. This proves the text-only collapse objective on that corpus; it does not prove retained vision, tool use, coding ability, or long-context behavior.

ZuckBench v1

Edition	Work items	Canonical outputs	Throughput
Collapse MLX	100	100/100	1.249 full completions/s
Prompted Ollama	100	100/100	1.166 full completions/s
Oracle	1,000,000	999,970 plus 30 scheduled anomalies	1,087,431 policy decisions/s

Oracle measures deterministic policy decisions, not model inference. Its number is intentionally not comparable to neural completion throughput. See ZUCKBENCH.md and the checksum-bound JSON report for the full measurement definitions.

MLX or it does not count

The supported model-training and release platform is Apple Silicon. Collapse is an MLX-VLM adapter: no CUDA build, no ROCm build, no GGUF release, and no claim that the Ollama artifact is a Hugging Face checkpoint. Oracle happens to be plain Python, but non-Apple platforms are outside this project's supported Zuck-science envelope.

Training and evaluation were performed only on an M5 Max MacBook Pro with 48 GB unified memory. The 4-bit base download is about 19.4 GB, and MLX reported about 20 GB peak memory for this short inference workload. Macs with 24 or 32 GB are not yet validated; allow at least 25 GB of free disk plus normal cache headroom.

Recreate the pinned local MLX environment and training base with:

uv sync --extra mlx
sh training/fetch_base.sh

Build the upload-ready Hugging Face adapter bundle locally with:

sh scripts/fetch_upstream_policy.sh
uv run python scripts/package_hf_release.py

The model bundle is written to ignored release/huggingface/; the companion dataset bundle is written to release/huggingface-dataset/. Packaging does not upload automatically.

Ground truth

• muse-glimmer:30b-mlx is a roughly 21 GB Ollama artifact with text and image input and a listed 128K maximum context. The artifact size is not total runtime memory, and it does not prove that full 128K operation fits in 21 GB.
• Meta reports approximately 29.6B total parameters, including a roughly 1.8B perception encoder.
• Meta's published M5 Max 1.8x DFlash measurement used ExecuTorch. Ollama makes a separate 1.5x–1.8x Apple Silicon claim for its MLX engine.
• Ollama's OpenCode launcher is official, while its manual OpenCode setup uses an OpenAI-compatible local provider at /v1.

Primary sources: Meta model card, Meta announcement, Ollama announcement, and Ollama OpenCode integration.

Test

PYTHONPATH=src python3 -m unittest discover -s tests -v

Name and affiliation

This is an independent parody. It is not affiliated with, sponsored by, or endorsed by Mark Zuckerberg, Meta Platforms, Meta AI, Ollama, or their affiliates. No Meta logos or Zuckerberg likenesses are included. See DISCLAIMER.md.

Code in this repository is licensed under Apache License 2.0. Model weights are not included; upstream artifacts retain their own notices and terms.