JANGQ-AI/Qwen3.5-27B-JANG_4S

CRITICAL FIX (2026-03-19): Fixed eos_token_id — previous versions caused infinite thinking loops. You MUST re-download this model if you downloaded before today.

Update (2026-03-18): Models updated to v2.1 with VLM support, proper tokenizer, and fixed configs. If you downloaded before this date, please re-download.

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]".

---

Qwen3.5-27B — JANG_4S (4-bit, 6-bit attention) — VLM

JANG — Jang Adaptive N-bit Grading | Mixed-Precision Quantization for Apple Silicon

     

JANG is fully open-source. Quantization engine, research, and full commit history: github.com/jjang-ai/jangq. Created by Jinho Jang.

Results (200-question MMLU)

Model	MMLU	Size	Speed
JANG_4S	84.5%	16 GB	35 tok/s
MLX 4-bit	84.5%	14 GB	20 tok/s

JANG_4S matches MLX 4-bit quality with 75% faster inference (35 vs 20 tok/s).

Per-Subject Scores (JANG_4S vs MLX 4-bit)

Subject	JANG_4S	MLX 4-bit
Abstract Algebra	15/20	15/20
Anatomy	17/20	17/20
Astronomy	19/20	19/20
College CS	16/20	16/20
College Physics	17/20	17/20
HS Biology	19/20	19/20
HS Chemistry	17/20	17/20
HS Mathematics	13/20	13/20
Logical Fallacies	18/20	18/20
World Religions	18/20	18/20
Total	169/200	169/200

Specs

Metric	Value
Source	Qwen3.5-27B
Architecture	Dense hybrid (GatedDeltaNet SSM + full attention)
Profile	JANG_4S (CRITICAL=6, IMPORTANT=4, COMPRESS=4)
Average bits	~4.15
GPU Memory	14.8 GB
Speed	35 tok/s
VLM	Yes (vision encoder preserved)
Format	v2 (MLX-native, instant load)

Install

pip install "jang[mlx]"

Quick Start

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/Qwen3.5-27B-JANG_4S")
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

VLM Inference

from jang_tools.loader import load_jang_vlm_model
from mlx_vlm import generate

model, processor = load_jang_vlm_model("JANGQ-AI/Qwen3.5-27B-JANG_4S")
prompt = processor.tokenizer.apply_chat_template(
    [{"role": "user", "content": [
        {"type": "image", "image": "photo.jpg"},
        {"type": "text", "text": "Describe this image."}
    ]}], add_generation_prompt=True, tokenize=False, enable_thinking=False)
result = generate(model, processor, prompt, ["photo.jpg"], max_tokens=200)
print(result.text)

Links

• GitHub | HuggingFace | MLX Studio | PyPI

---

한국어

Qwen3.5-27B — JANG_4S

JANG은 Apple Silicon을 위한 혼합정밀도 양자화 포맷입니다.

모델	MMLU	크기	속도
JANG_4S	84.5%	16 GB	35 tok/s
MLX 4-bit	84.5%	14 GB	20 tok/s

pip install "jang[mlx]"

GitHub · HuggingFace · MLX Studio

---

장진호 제작 · Created by Jinho Jang — jangq.ai · @dealignai