jang 2.5.28

JANG — Adaptive Mixed-Precision Quantization for Apple Silicon. v2.5.28: DSV4 JANGTQ conversion defaults to the V3 runtime-candidate lane with MTP dropped and explicit bit-plan metadata; std remains baseline-only. v2.5.27: ZAYA JANGTQ/MXFP converters and tools-only CCA capability contract.

MLX Studio — the only app that natively supports JANG models with reasoning

---

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

Highlights

• 397B on 128 GB Mac — JANG_1L: 112 GB, 36 tok/s, 86.5% MMLU with reasoning
• DeepSeek-V4-Flash runtime — first MLX runtime for DSV4's hybrid SWA + CSA + HSA attention (sliding-window + compressor-pool + hash-sparse) with mHC residuals and 1M-context pool quant cache
• JangPress for routed-MoE — load 167 GB Kimi-K2.6 on a 128 GB Mac via mmap + madvise cold-tier eviction (see docs/JANGPRESS.md)
• Nemotron-Cascade-2 in 10 GB — IMO Gold Medal reasoning model at 130 tok/s on 16 GB MacBooks
• MiniMax: only JANG works — MLX scores 25% (random), JANG scores 74%
• Nemotron-3-Super-120B in 43 GB — first working Nemotron-H quantization for Apple Silicon
• bfloat16 auto-detection — fixes float16 overflow on 512-expert models
• Reasoning mode — <think>...</think> with configurable thinking on/off

Results (200-question MMLU)

Qwen3.5-397B-A17B — JANG runs where MLX can't

Model	No-Think	Reasoning	Size	Speed
JANG_1L	81.0%	86.5%	112 GB	36 tok/s
JANG_2L	79.5%	92.0%	187 GB	36 tok/s
MLX 4-bit	81.5%	94.0%	209 GB	~36 tok/s
MLX 2/3-bit	NaN	NaN	—	—

MLX cannot quantize 397B below 4-bit (float16 overflow). JANG solves this with bfloat16.

Nemotron-Cascade-2-30B — IMO Gold Medal in 10 GB

Model	No-Think	Reasoning	Size	Speed
JANG_2L	59.0%	88.0%	10.3 GB	130 tok/s
JANG_4M	69.0%	93.0%	17 GB	55 tok/s
MLX 4-bit	69.0%	92.5%	16.6 GB	—
MLX 6-bit	71.0%	94.5%	23.9 GB	—

JANG_4M beats MLX 4-bit (93.0% vs 92.5%) at the same size.

Nemotron-3-Super-120B — Only JANG can go below 4-bit

Model	No-Think	Reasoning	Size	Speed
JANG_2L	75.0%	86.0%	43 GB	52 tok/s
JANG_4M	72.5%	93.0%	63 GB	55 tok/s
MLX 4-bit	71.0%	93.5%	63 GB	60 tok/s
MLX 3-bit	Crashes	—	—	—

MLX mlx_lm.convert crashes on Nemotron's mtp.* weights. Only JANG can produce sub-4-bit.

MiniMax-M2.5 — JANG is the ONLY working option

Model	MMLU	Size
JANG_2L	74%	63 GB
JANG_3M	74.5%	82 GB
MLX 4-bit	26.5%	120 GB
MLX 3-bit	24.5%	93 GB
MLX 2-bit	25%	—

MLX is broken on MiniMax at ALL bit levels (~25% = random). MiniMax has 256 experts — MLX compresses attention to the same bits as expert MLP, destroying coherence.

Qwen3.5 MoE (122B, 35B)

Model	JANG	MLX 4-bit	JANG Size	MLX Size
122B JANG_4K	86%	85%	69 GB	64 GB
122B JANG_2S	79%	56.5% (2-bit)	38 GB	36 GB
35B JANG_4K	77.5%	77.0%	16.7 GB	18 GB
35B JANG_2S	65.5%	~20% (2-bit)	12 GB	10 GB

The Full Picture: JANG vs MLX Across All Models

Model	JANG Best	MLX Best	JANG Size	MLX Size	MLX Broken?
Qwen3.5-397B	92.0%	94.0%	187 GB	209 GB	NaN below 4-bit
Qwen3.5-397B (128 GB Mac)	86.5%	—	112 GB	Can't fit	—
Nemotron-Cascade-2	93.0%	92.5%	17 GB	16.6 GB	—
Nemotron-Cascade-2 (16 GB Mac)	88.0%	—	10.3 GB	Can't fit	—
Nemotron-Super-120B	93.0%	93.5%	63 GB	63 GB	Crashes below 4-bit
Nemotron-Super-120B (64 GB Mac)	86.0%	—	43 GB	Can't fit	—
MiniMax-M2.5	74.5%	26.5%	82 GB	120 GB	Broken at ALL bits
Qwen3.5-122B	86%	85%	69 GB	64 GB	56.5% at 2-bit
Qwen3.5-35B	77.5%	77.0%	16.7 GB	18 GB	~20% at 2-bit

JANG wins at every size point. At equivalent sizes, JANG matches or beats MLX. At smaller sizes, JANG runs where MLX literally cannot (NaN, crashes, or random output).

Why MLX Fails on MoE Models

On MoE models, attention is only 1-5% of total parameters but controls 100% of coherence. MLX compresses everything equally:

MLX 4-bit: attention at 4-bit, experts at 4-bit → works but wastes bits on experts
MLX 2-bit: attention at 2-bit, experts at 2-bit → attention breaks → model breaks

JANG 2-bit: attention at 8-bit, experts at 2-bit → attention preserved → model works

The more experts a model has, the worse MLX performs at low bits:

• 128 experts (Cascade-2): MLX 4-bit still works, JANG slightly better
• 256 experts (122B, MiniMax): MLX 2-bit breaks badly, JANG dominates
• 512 experts (397B, Super-120B): MLX NaN/crash below 4-bit, only JANG works

DeepSeek-V4-Flash runtime — Hybrid SWA + CSA + HSA attention

DSV4-Flash is the first DeepSeek-class model whose attention is not plain MLA. JANG ships the only Apple-Silicon runtime that handles all three layer types correctly:

Layer type	Description	JANG path
SWA (Sliding-Window Attention)	local windowed attention; cheap fast layers	jang_tools.dsv4.mlx_model.DSV4SlidingWindow
CSA (Compressor / Compressed-State Attention)	per-layer compressor pool + windowed buffer; carries pooled-K/V across the whole context	DeepseekV4Cache + accumulate_windows + update_pool
HSA (Hash-Sparse Attention via Indexer)	indexer chooses top-K hash buckets per-token; sparse attention over the global compressed pool	Indexer module + DSV4LayerCache indexer state
mHC residuals	multi-Head-Compressed residual stream stabilises 43-layer 284B model	residual-axis fix in mlx_model.py
Pool quant cache	quantized streaming KV pool with RotatingKVCache.trim()-style proportional row trim (mirrors llama.cpp dsv4_clear_rows)	DeepseekV4Cache.trim(n)

End-to-end verified: 1M-context pool accumulation, multi-turn /v1/chat/completions with prefix-cache reuse, dual-mode reasoning (enable_thinking=true|false), 20.5 tok/s decode on M3 Ultra @ JANGTQ2 (79.5 GB), 24.5 tok/s on Swift.

# Convert from FP8 source bundle (~50 min on M3 Ultra)
jang-convert-dsv4-jangtq /path/to/DeepSeek-V4-Flash --out ~/models/DSV4-Flash-JANGTQ

# Inference (Python)
python -m jang_tools.load_jangtq ~/models/DSV4-Flash-JANGTQ

For runtime traps (chat-template injection, repetition collapse with enable_thinking=ON, rope_parameters vs rope_scaling, mandatory tq_bits strip on Swift), see the in-package jang_tools/dsv4/README.md.

JangPress — fit routed-MoE bundles bigger than RAM

Routed-MoE bundles like Kimi-K2.6 (167 GB), MiniMax-M2.7, and DSV4-Flash JANGTQ would normally OOM on a 128 GB Mac. JangPress is the load-time memory policy in osaurus-ai/vmlx-swift-lm that combines mmap-backed safetensors, per-token router-aware MADV_DONTNEED over canonical routed-expert pages, and an optional prestack overlay to keep post-load RSS around 1 GB on 128 GB hosts.

This package ships the JANG-side glue:

# Serve a 167 GB Kimi bundle on a 128 GB Mac
cd scripts/jangpress
./kimi_serve.sh ~/.mlxstudio/models/JANGQ-AI/Kimi-K2.6-Med-JANGTQ 100 8082
./kimi_mmlu.sh Kimi-K2.6-Med-JANGTQ chat 8082

Full guide: docs/JANGPRESS.md.

Install

pip install "jang[mlx]>=2.5.18"

For Vision-Language models:

pip install "jang[vlm]>=2.5.18"

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_2L

Run inference

from jang_tools.loader import load_jang_model
from mlx_lm import generate

model, tokenizer = load_jang_model("JANGQ-AI/Qwen3.5-397B-A17B-JANG_1L")

# With reasoning (recommended for hard questions)
messages = [{"role": "user", "content": "Prove that sqrt(2) is irrational."}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False,
    add_generation_prompt=True, enable_thinking=True)
result = generate(model, tokenizer, prompt=prompt, max_tokens=2048)

# Without reasoning (faster)
prompt = tokenizer.apply_chat_template(messages, tokenize=False,
    add_generation_prompt=True, enable_thinking=False)
result = generate(model, tokenizer, prompt=prompt, max_tokens=100)

VLM (Vision-Language) inference

from jang_tools.loader import load_jang_vlm_model
from mlx_vlm import generate as vlm_generate

model, processor = load_jang_vlm_model("JANGQ-AI/Qwen3.5-397B-A17B-JANG_2L")
messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": "Describe this image."},
]}]
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
result = vlm_generate(model, processor, prompt=prompt, image=["photo.jpg"], max_tokens=200)

MMLU Benchmark

python -m jang_tools.benchmark /path/to/model --max-thinking 1024

Smart two-pass: no-thinking first, then reasoning retry on wrong answers. Checkpointing, forced answers, full output logging.

Pre-quantized Models

Model	Profile	MMLU	Size	Fits
Qwen3.5-397B JANG_1L	2.1-bit	86.5%*	112 GB	128 GB Mac
Qwen3.5-397B JANG_2L	3.7-bit	92.0%*	187 GB	256 GB Mac
Nemotron-Cascade-2 JANG_2L	2.3-bit	88.0%*	10 GB	16 GB Mac
Nemotron-Cascade-2 JANG_4M	4.1-bit	93.0%*	17 GB	24 GB Mac
Nemotron-Super-120B JANG_2L	2.8-bit	86.0%*	43 GB	64 GB Mac
Nemotron-Super-120B JANG_4M	4.1-bit	93.0%*	63 GB	64 GB Mac
Qwen3.5-122B JANG_4K	4.0-bit	86%	69 GB	192 GB Mac
Qwen3.5-122B JANG_2S	2.1-bit	79%	38 GB	64 GB Mac
Qwen3.5-35B JANG_4K	4.0-bit	77.5%	17 GB	36 GB Mac
MiniMax-M2.5 JANG_2L	2.3-bit	74%	63 GB	128 GB Mac
Qwen3.5-27B JANG_4S	4.1-bit	84.5%	16 GB	24 GB Mac

* with reasoning mode

Full collection

Profiles

Profile	Type	Bits	Best for
JANG_4K	K-quant	4.0	Same size as MLX 4-bit, smarter
JANG_4M	Profile	4.0	8-bit attention, 4-bit experts
JANG_4S	Profile	4.0	Dense models (27B)
JANG_3K	K-quant	3.0	Same size as MLX 3-bit, smarter
JANG_2L	Profile	~2.3	Quality 2-bit, best for MoE
JANG_1L	Profile	~2.1	Maximum quality 2-bit

App Developers: Add JANG Support

JANG models are standard MLX safetensors. If your app loads MLX quantized models, adding JANG is minimal work.

Quickest Integration (5 lines)

# Detect JANG model
from pathlib import Path
is_jang = (Path(model_path) / "jang_config.json").exists()

# Load with jang-tools
if is_jang:
    from jang_tools.loader import load_jang_model
    model, tokenizer = load_jang_model(model_path)
    # model is a standard mlx_lm model — use like any MLX model

What's Different from Standard MLX

1. Mixed bit widths — different tensors have different bits (attention at 8-bit, experts at 2-bit). Each QuantizedLinear needs its bits and group_size set from tensor shapes.
2. bfloat16 for large models — 512+ expert models need model.set_dtype(mx.bfloat16) to prevent float16 overflow.
3. Nemotron-H weight renaming — switch_mlp.up_proj→fc1, down_proj→fc2, gate dequantization.

Full Integration Guide

See INTEGRATION.md for complete step-by-step with code for:

• Loading without jang-tools dependency
• Per-tensor bit inference from shapes
• bfloat16 auto-detection
• Nemotron-H special handling
• Chat template with thinking on/off
• VLM support
• Edge cases and gotchas

Supported Architectures

• Qwen3.5 (hybrid SSM + MoE + VLM) — 4B, 9B, 27B, 35B, 122B, 397B
• Nemotron-H (Mamba-2 + Latent MoE + Attention) — Cascade-2 30B, Super-120B
• MiniMax-M2.5 (256-expert MoE, FP8 source)
• DeepSeek-V2/V3 (MLA + MoE)
• Mixtral / Qwen2-MoE (standard MoE)
• Dense Transformers (Llama, Mistral, Gemma, Phi)
• Vision-Language (Qwen3.5-VL, Pixtral)
• Mamba / Hybrid SSM (Jamba, Nemotron-H)
• FP8 source models (auto-dequantization)
• Mistral Small 4 (119B MoE + MLA + Pixtral VL) — coming soon

Changelog

v2.1.5 (2026-03-21)

• Nemotron-H loader: fc1/fc2 rename, gate weight dequantization, mtp.* key filtering
• bfloat16 auto-detection for 512+ expert models (prevents float16 overflow)
• MLP asymmetry floors: gate_proj=4-bit, down_proj=3-bit for 512+ expert models
• Benchmark script: smart two-pass MMLU with reasoning, checkpointing, forced answers
• eos_token_id auto-fix for Qwen3.5 (248044→248046)
• Auto-copy all .py files for trust_remote_code models
• Nemotron-3-Super-120B: 86% MMLU at 43 GB
• Qwen3.5-397B: 92% MMLU at 187 GB, 86.5% at 112 GB

v2.1.4 (2026-03-19)

• MLP asymmetry fix for 512-expert models
• eos_token_id auto-fix for Qwen3.5
• Auto-copy custom .py files

v2.1.3 (2026-03-18)

• Per-tensor group_size (router=64, experts=128 for 150+ expert models)
• Precision floor rules for shared expert
• VLM support for all Qwen3.5 models

How It Works

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

CRITICAL  (attention, MoE routers, MLA latent)  →  6-8 bit  →  Controls coherence
IMPORTANT (embeddings, linear attention)         →  4-6 bit  →  Moderate sensitivity
COMPRESS  (MLP, MoE experts)                     →  2-4 bit  →  95%+ of parameters

On MoE models, attention is only 1-5% of parameters. Boosting it to 8-bit costs ~2% overhead but dramatically improves quality. MLX compresses everything equally — that's why it breaks on MoE models at low bits.

Technical Features

• bfloat16 compute: Auto-detected for 512+ expert models. Prevents float16 overflow at shared expert down_proj.
• MLP asymmetry: gate_proj gets 4-bit floor (SiLU amplifier), down_proj gets 3-bit floor for 512+ expert models.
• FP8 dequantization: Handles FP8 source models (MiniMax, Nemotron) automatically.
• Latent MoE: Supports Nemotron-H's fc1/fc2_latent_proj compression.
• v2 format: MLX-native safetensors, instant mmap loading, no repack needed.

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 (inference), mlx-vlm>=0.1 (VLM)

Links

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

---

한국어

JANG은 Apple Silicon을 위한 혼합정밀도 양자화 포맷입니다. MLX를 위한 GGUF.

모델	MMLU	크기	최소 Mac
Qwen3.5-397B JANG_1L	86.5%*	112 GB	128 GB
Nemotron-Cascade-2 JANG_2L	88.0%*	10 GB	16 GB
Nemotron-Super-120B JANG_2L	86.0%*	43 GB	64 GB
MiniMax-M2.5 JANG_2L	74%	63 GB	128 GB

* 추론 모드 사용

pip install "jang[mlx]>=2.5.18"

GitHub · HuggingFace · MLX Studio · PyPI

---

장진호 제작 · Created by Jinho Jang — jangq.ai