jang 2.0.0

JANG — Adaptive Mixed-Precision Quantization for Apple Silicon. The GGUF equivalent for MLX.

MLX Studio — the only app that natively supports JANG models

---

Early Adoption: LM Studio, Ollama, oMLX, Inferencer do not support JANG yet. Use MLX Studio or pip install "jang[mlx]". Ask your favorite app's creators to add JANG support!

---

Jang Adaptive N-bit Grading

Mixed-Precision Quantization for Apple Silicon

The GGUF equivalent for MLX — models stay quantized in GPU memory at full Metal speed.

Website • Models • PyPI • Format Spec

Results (200-question MMLU)

MoE at 4-bit: JANG_4K beats MLX

Model	JANG_4K	MLX 4-bit	JANG Size	MLX Size
Qwen3.5-122B	86%	85%	69 GB	64 GB
Qwen3.5-35B	77.5%	75.5%	16.7 GB	18 GB

MoE at 2-bit: JANG dominates

Model	JANG_2S	MLX 2-bit	JANG Size	MLX Size
Qwen3.5-122B	79%	56.5%	38 GB	36 GB
Qwen3.5-35B	65.5%	~20%	12 GB	10 GB

MiniMax: JANG is the ONLY working option

Model	JANG_2L	MLX 4-bit	MLX 3-bit	MLX 2-bit
MiniMax-M2.5	74%	26.5%	24.5%	25%

MLX is broken on MiniMax at ALL bit levels (~25% = random). JANG scores 74%.

Install

pip install jang

For inference on Apple Silicon:

pip install "jang[mlx]"

For Vision-Language models:

pip install "jang[vlm]"

Quick Start

Convert any model

# K-quant 4-bit (same size as MLX, smarter allocation)
jang convert Qwen/Qwen3.5-35B-A3B -p 4

# 2-bit for extreme compression
jang convert Qwen/Qwen3.5-122B-A10B -p 2

# Specific profile
jang convert model -p JANG_2S

Run inference

from jang_tools.loader import load_jang_model
from mlx_lm.sample_utils import make_sampler
from mlx_lm.generate import generate_step
import mlx.core as mx

model, tokenizer = load_jang_model("JANGQ-AI/Qwen3.5-122B-A10B-JANG_2S")
sampler = make_sampler(temp=0.7)

tokens = tokenizer.encode("What is photosynthesis?")
for tok, _ in generate_step(prompt=mx.array(tokens), model=model, max_tokens=200, sampler=sampler):
    t = tok.item() if hasattr(tok, 'item') else int(tok)
    print(tokenizer.decode([t]), end="", flush=True)
    if t == tokenizer.eos_token_id:
        break

Upgrade v1 models to v2 (instant loading)

jang upgrade /path/to/model

CLI Commands

Command	Description
jang convert <model> -p <profile>	Convert HuggingFace model to JANG
jang upgrade <model>	Upgrade v1 model to v2 (instant load)
jang inspect <model>	Show bit allocation and model info
jang validate <model>	Validate a JANG model directory
jang estimate <params>	Estimate sizes (e.g., jang estimate 122B)

v2 Format — Instant Loading

JANG v2 stores weights in MLX-native format. Like GGUF — the file IS the runtime format. No conversion at load time.

	v2 (current)	v1 (legacy)
Load time	Seconds (mmap)	5-10 minutes (repack)
File size	Same	Same

New conversions automatically use v2. Existing v1 models can be upgraded with jang upgrade.

Profiles

Profile	Type	Bits	Best for
JANG_4K	K-quant	4.0	Same size as MLX 4-bit, smarter
JANG_3K	K-quant	3.0	Same size as MLX 3-bit, smarter
JANG_2S	Profile	~2.1	Tightest 2-bit, near MLX 2-bit size
JANG_2L	Profile	~2.3	Quality 2-bit
JANG_1L	Profile	~2.2	Maximum quality 2-bit

Pre-quantized Models

Model	Profile	MMLU (200q)	HumanEval	Size
Qwen3.5-122B-A10B	JANG_4K	86%	95%	69 GB
Qwen3.5-122B-A10B	JANG_2S	79%	90%	38 GB
Qwen3.5-35B-A3B	JANG_4K	77.5%	90%	16.7 GB
Qwen3.5-35B-A3B	JANG_2S	65.5%	—	12 GB
MiniMax-M2.5	JANG_2L	74%	—	89 GB

Supported Architectures

Dense Transformer, Mixture of Experts, Hybrid SSM, Linear Attention (GatedDeltaNet), MLA (DeepSeek), Vision-Language, Mamba, FP8 source models (MiniMax, DeepSeek).

How It Works

JANG redistributes bits based on tensor sensitivity — same total size, smarter allocation:

CRITICAL  (attention, MoE routers)   →  6-8 bit  →  Controls coherence
IMPORTANT (embeddings, linear attn)  →  4-6 bit  →  Moderate sensitivity
COMPRESS  (MLP, MoE experts)         →  2-4 bit  →  98% of parameters

K-quant profiles (JANG_4K, JANG_3K) redistribute within the same bit budget — boost attention, compensate with least-important MLP. Same size as MLX, smarter allocation. Like GGUF K-quants.

Requirements

• Python: 3.11+
• Conversion: any platform (numpy + safetensors)
• Inference: Apple Silicon Mac (M1/M2/M3/M4) with MLX
• Dependencies: safetensors>=0.4, numpy>=1.24, tqdm>=4.60, huggingface_hub>=0.20
• Optional: mlx>=0.22, mlx-lm>=0.20 (for inference), mlx-vlm>=0.1 (for VLM)

Links

• GitHub | HuggingFace | MLX Studio | PyPI | Format Spec

---

한국어

JANG이란?

JANG은 Apple Silicon을 위한 오픈소스 혼합정밀도 양자화 포맷입니다. MLX를 위한 GGUF와 같은 역할을 합니다.

결과 (200문항 MMLU)

4-bit: JANG_4K가 MLX 4-bit보다 우수 (MoE 모델)

모델	JANG_4K	MLX 4-bit	크기
Qwen3.5-122B	86%	85%	69 vs 64 GB
Qwen3.5-35B	77.5%	75.5%	16.7 vs 18 GB

2-bit: JANG이 MLX를 압도

모델	JANG_2S	MLX 2-bit	크기
Qwen3.5-122B	79%	56.5%	38 vs 36 GB
Qwen3.5-35B	65.5%	~20%	12 vs 10 GB

MiniMax: JANG만 작동

모델	JANG_2L	MLX 4-bit	MLX 3-bit	MLX 2-bit
MiniMax-M2.5	74%	26.5%	24.5%	25%

설치

pip install "jang[mlx]"

호환성

현재 **MLX Studio**만 JANG 포맷을 기본 지원합니다. LM Studio, Ollama, oMLX, Inferencer 등은 아직 지원하지 않습니다. 좋아하는 앱의 개발자에게 JANG 지원을 요청해 주세요!

GitHub · HuggingFace · MLX Studio · PyPI

---

장진호 제작 · Created by Jinho Jang — jangq.ai