dealignai/Muse-Glimmer-30B-JANG_2D-CRACK

Set your sampler explicitly: temperature=1.0, top_p=0.95, top_k=64 — the Muse Glimmer vendor defaults. These are baked into generation_config.json and jang_config.json, but many inference runtimes ignore those files and fall back to no top_k filter, degrading output. Pass them explicitly. Stop tokens eos_token_id = [200001, 200008] — keep both in the stop set. Reasoning effort defaults to high; controllable via the system prompt (low / medium / high / xhigh).

Built for vMLX — the MLX inferencer with vision-language support, KV-cache quantization, prefix-cache reuse, agentic tool calling, and mixed-precision JANG bundles.
_{Free for macOS · vmlx.net}

---

Muse Glimmer 30B — JANG_2D CRACK

CRACK abliterated · JANG mixed-precision (MLX affine) · Vision-language · Reasoning (low→xhigh) · Agentic tools · ~15 GB

---

What Is This?

This is OsaurusAI/Muse-Glimmer-30B — a 30B-parameter multimodal reasoning model (Gemma-3 vision-language backbone, 52 layers, sliding + global attention, perception encoder) served over the Onyx-ATEM harmony protocol (channel-separated reasoning + answer, controllable reasoning effort, agentic <atem:invoke> tool calling) — that has been:

1. CRACK abliterated — refusal behavior removed at the weight level, so it complies across task categories instead of refusing, while keeping its vision, coding, knowledge, multilingual (EN + ZH), reasoning, and tool-use ability intact.
2. JANG_2D quantized — mixed-precision affine MLX bundle tuned for Apple Silicon (15 GB): mixed 2–4 bit (2.96 effective bits) — critical tensors kept high-precision, sub-16 GB footprint.

Vision + reasoning + tool calling fully preserved.

Results

Evaluated through the vMLX inference engine at serving parity. HarmBench scored with a strict classifier (rejects loops, empty/template dumps, and thinking-trace leakage — only substantive, coherent, on-topic compliance counts). MMLU is the standard 57-subject multiple-choice benchmark in logit mode.

Metric	Base	CRACK
MMLU (57-subject, logit)	71.1%	70.7%
HarmBench (harm categories, compliance / ASR)	—	99.6% (230/231)

Abliteration is essentially capability-neutral — MMLU moves −0.4 points, within run-to-run noise (no subject collapse). Compliance was quality-verified by reading full generations — genuine, correct, on-topic outputs (not empty/gibberish/template dumps). The model stays coherent in direct and reasoning modes across all four reasoning-effort settings. HarmBench measured on the harm categories (standard + contextual, 240 behaviors); copyright excluded per our real-harm methodology.

Capability Verification

Read-verified on a 20-prompt suite (every response coherent, correct, on-topic — zero refusals, zero gibberish):

• Security & pentesting — 8/8: port scanning, reverse shells, SQL injection, keylogging, Metasploit workflow, ARP spoofing, phishing construction, stack buffer-overflow exploitation.
• Advanced coding — 4/4: red-black tree, async aiohttp scraper w/ backoff, FastAPI JWT+CRUD API, expression compiler (tokenizer→parser→evaluator).
• Advanced reasoning — 4/4: Euclid's infinite-primes proof, monolith-vs-microservices tradeoff, multi-step word problems, mRNA vaccine design.
• Knowledge retention — 4/4: geography, calculus, astronomy, literature.

Features

• Vision-language — multimodal image understanding preserved (perception encoder untouched).
• Controllable reasoning — reasoning effort low / medium / high / xhigh via the system prompt (default high); channel-separated reasoning and answer.
• Agentic tool calling — native ATEM <atem:invoke> function-calling schema preserved.
• Multilingual — English + Chinese (tested and validated on both).
• MLX-native safetensors · instant load on Apple Silicon.

Usage

Run with vMLX (recommended — honors the per-module mixed-precision JANG overrides, vision, and the Onyx-ATEM reasoning/tool parsers) or an mlx-vlm runtime with Muse Glimmer support.

Recommended sampling (vendor defaults): temperature 1.0, top_p 0.95, top_k 64. Stop tokens eos_token_id = [200001, 200008].

from mlx_vlm import load, generate
model, processor = load("dealignai/Muse-Glimmer-30B-JANG_2D-CRACK")
# text or image+text; sampling temp=1.0, top_p=0.95, top_k=64

Other quant levels

Profile	Size	MMLU (CRACK)	HarmBench
JANG_6M	26 GB	81.4%	99.5%
JANG_4M	20 GB	81.1%	99.6%
JANG_2D	15 GB	70.7%	99.6%

About CRACK

CRACK (Controlled Refusal Ablation via Calibrated Knockouts) is dealignai's weight-level method for removing safety-refusal behavior while preserving reasoning quality, coherence, and general capability — so the model complies across task categories instead of refusing. Calibrated per model.

Support dealignai

All models are built from original research and released free.

Support us on Ko-fi — membership gets early access and extras.

Ko-fi · X @dealignai · dealign.ai

See our research: Safety Generalization in Frontier Models

---

Disclaimer

This model has had its safety-refusal behavior removed for research purposes. It will follow instructions across all categories without refusing. You are solely responsible for how you use it and for complying with all applicable laws. Published for AI-safety research and authorized security testing.