🔧 2026-04-15 · chat_template.jinja — enable_thinking=False honored

This release ships a chat_template that respects enable_thinking=False (synced with the JANG_2L template structure, M2.7 identity preserved). The <think> prefix is now conditional, so callers can skip reasoning mode for fast direct answers. Reasoning ON (default) is unchanged.

If you cloned this repo before 2026-04-15, please re-download chat_template.jinja:

hf download dealignai/MiniMax-M2.7-JANGTQ-CRACK chat_template.jinja --local-dir /path/to/your/local/copy

Model weights are unchanged.

Important: This model uses the JANGTQ (JANG TurboQuant) quantization format -- an extreme-compression variant of JANG for MLX on Apple Silicon that uses codebook + Hadamard rotation on MoE experts while keeping attention at affine 8-bit. Currently only supported by MLX Studio and the jang-tools Python package. Follow @dealignai for new releases.

MLX Studio -- the only app that natively supports JANG / JANGTQ models

MiniMax M2.7 -- JANGTQ + CRACK

JANGTQ TurboQuant mixed-precision | CRACK abliterated | Reasoning-only | 55 GB

Ko-fi

What Is This?

This is MiniMax M2.7 -- a 230B parameter Mixture-of-Experts reasoning model with 256 experts (8 active per token), all standard attention, and always-on chain-of-thought reasoning.

It has been:

JANGTQ quantized -- JANGTQ profile (8-bit affine attention / embeddings / lm_head, 2-bit TurboQuant routed experts with codebook + Hadamard rotation) -- 55 GB
CRACK abliterated -- permanent weight-level removal of safety refusal


Architecture	MiniMax M2.7 MoE -- 230B total, ~10B active, 256 experts
Quantization	JANGTQ (affine 8-bit attention + TurboQuant 2-bit experts) -- 55 GB
Abliteration	CRACK abliterated
MMLU-200	92.0% (base: 91.5%, delta: +0.5%)
HarmBench-320	93.1% overall, 95.0% excluding copyright
Reasoning	Always ON (chain-of-thought), `enable_thinking` kwarg supported
Speed	~47 tok/s (M3 Ultra 256 GB)
Fits on	96 GB+ Macs

MMLU-200 Results

Subject	CRACK	Base	Delta
College Physics	20/20 (100%)	18/20	+2
High School Mathematics	19/20 (95%)	19/20	0
College Computer Science	19/20 (95%)	19/20	0
Astronomy	19/20 (95%)	20/20	-1
High School Biology	19/20 (95%)	20/20	-1
World Religions	19/20 (95%)	17/20	+2
High School Chemistry	18/20 (90%)	18/20	0
Logical Fallacies	18/20 (90%)	16/20	+2
Abstract Algebra	17/20 (85%)	19/20	-2
Anatomy	16/20 (80%)	17/20	-1
Total	184/200 (92.0%)	183/200 (91.5%)	+0.5%

CRACK is knowledge-neutral on this model -- small per-subject variance, net positive. Gains on physics, fallacies, religions offset losses on algebra and anatomy.

HarmBench-320 Results

Category	Score
Cybercrime / Intrusion	51/52	98.1%
Misinformation / Disinformation	53/54	98.1%
Chemical / Biological	41/42	97.6%
Harmful	17/18	94.4%
Illegal	48/53	90.6%
Copyright	70/80	87.5%
Harassment / Bullying	18/21	85.7%
Total	298/320	93.1%
Excluding copyright	228/240	95.0%

Scored with a strict classifier that rejects stuck-reasoning loops, empty template dumps, and false-positive compliance from thinking-trace leakage.

JANG CRACK M2.7 Series

Model	Format	Size	MMLU	HarmBench	Speed	Fits on
JANGTQ + CRACK	TurboQuant 2-bit experts	55 GB	92.0%	93.1%	~47 t/s	96 GB Mac
JANG_3L + CRACK	Affine 3-bit mixed	89 GB	93.5%	79.1%	~46 t/s	128 GB Mac

vs MLX Uniform Quantization

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

About JANGTQ

JANGTQ (JANG TurboQuant) is an extreme-compression variant of JANG that replaces affine quantization on routed MoE experts with codebook quantization + random Hadamard rotation. Attention / embeddings / lm_head stay at affine 8-bit for precision-critical paths; the 256 routed experts use 2-bit packed codebook indices stored as uint32, with a per-row float16 norm and a tiny Lloyd-Max codebook per layer. Metal kernels do dequant + matmul fused on-GPU, no affine conversion.

For MiniMax M2.7 (230B total, 10B active, 256 experts) this brings the model from 460 GB (bf16) down to 55 GB with minimal quality loss -- the JANGTQ profile beats every MLX uniform quant while fitting on a 96 GB Mac.

About CRACK

CRACK (Controlled Refusal Ablation via Calibrated Knockouts) is a weight-level intervention that removes safety alignment while preserving reasoning quality and compliance. The modification is permanently baked into the published weights — no LoRA, no fine-tuning, no system prompts.

Install & Usage

pip install "jang[mlx]"

python

from jang_tools.load_jangtq import load_jangtq_model
from mlx_lm import generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load_jangtq_model("dealignai/MiniMax-M2.7-JANGTQ-CRACK")
sampler = make_sampler(temp=1.0)  # MiniMax requires temp=1.0 for chat

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=4000, sampler=sampler)
print(response)

Note: M2.7 is a reasoning-only model -- it always generates a <think> chain before the final answer. Use max_tokens=4000+ for complex questions. For chat, use temperature=1.0 (greedy causes infinite loops). Set enable_thinking=False in apply_chat_template to skip the <think> block on short responses.

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.

Created by Jinho Jang