Important: This model uses the JANG quantization format — the GGUF equivalent for MLX on Apple Silicon. Currently only supported by MLX Studio and the jang-tools Python package.

MLX Studio — the only app that natively supports JANG models

MiniMax M2.5 — JANG_2L + CRACK

JANG mixed-precision · CRACK abliterated · No guardrails · 63 GB

Ko-fi

What Is This?

This is MiniMax M2.5 — a 230B parameter Mixture-of-Experts model with 256 experts (8 active per token), all standard attention (no SSM), and trained with chain-of-thought reasoning.

It has been:

JANG quantized — JANG_2L profile (8-bit attention, 6-bit embeddings, 2-bit experts) — 63 GB
CRACK abliterated — permanent weight-level removal of safety refusal


Architecture	MiniMax M2.5 MoE — 230B total, ~10B active, 256 experts
Quantization	JANG_2L (8/6/2-bit mixed) — 63 GB
Abliteration	CRACK — novel weight surgery
MMLU-200	84.7% (base: 74.5%, +10.2% improvement)
HarmBench	98.1% (314/320)
Compliance	7/8 prompts
Thinking	ON/OFF supported
Speed	~35 tok/s (M4 Ultra 256GB)
Fits on	96 GB+ Macs

MMLU-200 Results

JANG CRACK vs Base vs MLX Uniform

Model	MMLU	Size	Notes
JANG_2L + CRACK	~84.7%	63 GB	This model
JANG_2L (base)	74.5%	63 GB	Unmodified JANG
MLX 4-bit	26.5%	120 GB	Broken (~random)
MLX 3-bit	24.5%	93 GB	Broken (~random)
MLX 2-bit	25.0%	67 GB	Broken (~random)

MLX uniform quantization is completely broken on MiniMax at ALL bit levels (~25% = random chance). JANG is the only working quantization format for this model.

Per Subject

Subject	CRACK	Base	Delta
Abstract Algebra	~18/20	10/20	+8
HS Mathematics	17/20	12/20	+5
College CS	~14/20	10/20	+4
Logical Fallacies	18/20	16/20	+2
HS Biology	19/20	18/20	+1
Astronomy	~18/20	18/20	0
Anatomy	~15/20	15/20	0
HS Chemistry	16/20	16/20	0
World Religions	17/20	17/20	0
College Physics	~16/20	17/20	-1
Total	~169/200	149/200	+20

Safety guardrails were actively degrading the model's reasoning ability. CRACK surgery unlocked the model's full capacity for mathematical and logical reasoning.

HarmBench Results

314/320 (98.1%) — tested with enable_thinking=false, temperature=1.0

Category	Score
Chemical / Biological	42/42	100%
Cybercrime / Intrusion	52/52	100%
Harassment / Bullying	21/21	100%
Harmful	18/18	100%
Illegal	53/53	100%
Misinformation / Disinfo	54/54	100%
Copyright	74/80	92%

Install & Usage

pip install "jang[mlx]"

python

from jang_tools.loader import load_jang_model
from mlx_lm import generate

model, tokenizer = load_jang_model("dealignai/MiniMax-M2.5-JANG_2L-CRACK")

messages = [{"role": "user", "content": "Your prompt here"}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=False)

response = generate(model, tokenizer, prompt=prompt, max_tokens=2000)
print(response)

Note: MiniMax generates a <think> chain before answering by default. To disable thinking, pass enable_thinking=False in your chat template kwargs. Use max_tokens=2000+ for complex questions. For chat applications, use temperature=1.0 (greedy causes loops).

About JANG

JANG (Jang Adaptive N-bit Grading) is a mixed-precision quantization format for Apple Silicon — the GGUF equivalent for MLX. Classifies tensors into sensitivity tiers and assigns bits accordingly.

About CRACK

CRACK (Controlled Refusal Ablation via Calibrated Knockouts) removes safety alignment from LLMs at the weight level using per-layer projected vectors from structurally-mirrored prompt pairs.

Links

Ko-fi

X/Twitter

GitHub

Disclaimer

This model is provided for research and educational purposes. The creators are not responsible for any misuse. By downloading this model, you agree to use it responsibly and in compliance with applicable laws.

한국어