Modelpublic

JANGQ-AI/MiniMax-M2.7-JANG_6M

sourceHugging Faceotherupdated 15d agoView on Hugging Face

1likes116downloads

Model Card

<h3 align="center">⚡ All JANG models are meant to be run in <a href="https://vmlx.net">vMLX</a></h3>

MiniMax-M2.7 JANG_6M

MiniMax M2.7 228B MoE — 6.03-bit mixed precision, 167 GB

Near-lossless quantization for maximum quality on Apple Silicon. </div>

Recommended: Run in [MLX Studio](https://mlxstudio.com) for best experience including thinking mode support and optimized MoE inference.

Important Settings

MiniMax M2.7 is an always-reasoning model. It thinks before answering on every prompt.

Setting	Value	Notes
Temperature	1.0	REQUIRED — greedy/temp=0 causes infinite thinking loops
Top P	0.95
Top K	40
Repetition Penalty	1.1	Optional, helps prevent loops

Model Details

Metric	Value
Source	`MiniMaxAI/MiniMax-M2.7` (FP8 E4M3)
Architecture	MoE (256 experts, top-8 active), GQA (48 heads / 8 KV), partial RoPE
Total Parameters	228.7B
Active Parameters	~1.4B per token
Profile	JANG_6M (CRITICAL=8-bit, IMPORTANT=6-bit, COMPRESS=6-bit)
Actual avg bits	6.03
Model size	167 GB
Format	JANG v2 (MLX-native safetensors, instant load)
group_size	128 (speed-optimized for 256 experts)
Routing	Sigmoid + bias correction (not softmax)
QK-norm	Full vector RMSNorm
Context	192K tokens

JANG_6M Bit Allocation

Tier	Components	Bits
CRITICAL	Attention (Q/K/V/O), lm_head	8
IMPORTANT	Embeddings	6
COMPRESS	Expert MLP (w1/w2/w3) — 98%+ of params	6
Passthrough	MoE router/gate (float16), norms, QK-norms	16

JANG protects routing and attention at full precision while compressing the 256 expert MLPs — where MoE models are most tolerant of quantization. The router is kept at float16 (no quantization) for maximum routing precision.

MMLU Benchmarks (200q, 10 subjects, reasoning ON)

Coming soon — benchmarks in progress.

Why JANG for MiniMax

Standard MLX quantization on MiniMax produces completely broken output at ALL bit levels (~25% MMLU = random guessing). JANG's mixed-precision approach is the only working quantized MiniMax on Apple Silicon.

On M2.5, JANG_2L achieved 74% MMLU vs MLX's 25% (random). M2.7 results pending.

All Quantizations

Model	Profile	Size	Avg Bits
JANG_2L	(8, 6, 2)	63 GB	2.10
JANG_3L	(8, 4, 3)	89 GB	3.08
JANG_4M	(8, 4, 4)	115 GB	4.06
JANG_6M	(8, 6, 6)	167 GB	6.03

Requirements

—Apple Silicon Mac with 192 GB unified memory
—MLX framework
—MLX Studio recommended

Tool Use / Agent Mode

MiniMax M2.7 uses interleaved thinking + tool calls — it reasons inside <think> blocks, then emits tool calls in <minimax:tool_call> format. Some clients (Opencode, etc.) may strip the <think> block and miss the tool call.

For tool-use clients, set enable_thinking=False in the chat template:

python

text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True,
    enable_thinking=False  # skips <think> injection for tool-use
)

MiniMax tool call format:

xml

<minimax:tool_call>
<invoke name="tool_name">
<parameter name="param1">value1</parameter>
</invoke>
</minimax:tool_call>

Usage

python

from jang_tools.loader import load_jang_model
from mlx_lm import generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load_jang_model("JANGQ-AI/MiniMax-M2.7-JANG_6M")
sampler = make_sampler(temp=1.0, top_p=0.95)

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "What is photosynthesis?"}],
    tokenize=False, add_generation_prompt=True
)
output = generate(model, tokenizer, prompt=prompt, max_tokens=2048, sampler=sampler)
print(output)

Support

MLX Studio | JANGQ | X @dealignai

Quantized by Jinho Jang (eric@jangq.ai) using JANG Tools v2.4.1.

This model is provided for research and personal use. Users are responsible for ensuring their use complies with applicable laws and the MiniMax license.