OsaurusAI/Qwen3.5-122B-A10B-JANG_2S

Qwen 3.5 122B-A10B — JANG_2S (Mixed-Precision, 2-bit)

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

     

---

Osaurus natively supports JANG models. Download at osaurus.ai.

---

Model Details

Property	Value
Base Model	Qwen 3.5 VL 122B-A10B
Architecture	MoE Transformer + Vision
Total Parameters	122B (10B active per token)
Profile	JANG_2S
Avg Bits/Weight	2.11
Bit Widths Used	2, 4, 6
Model Size	30.7 GB
Vision	Yes
Format	JANG v2 (MLX-native safetensors)

Benchmarks

200-question MMLU (20 per subject x 10 subjects). Thinking OFF (enable_thinking=False), greedy decoding (temp=0.0).

Model	MMLU	Size
JANG_2S (this)	79%	30.7 GB
MLX 2-bit	56.5%	36 GB
JANG_4K	86%	57.4 GB
MLX 4-bit	85%	64 GB

JANG_2S scores 79% vs MLX 2-bit's 56.5% — a +22.5 point improvement while being smaller (30.7 GB vs 36 GB). Fits on 48 GB Macs where MLX 4-bit cannot.

Per-Subject Breakdown

Subject	JANG_2S
Abstract Algebra	9/20
Anatomy	18/20
Astronomy	20/20
College CS	14/20
College Physics	15/20
HS Biology	19/20
HS Chemistry	18/20
HS Mathematics	11/20
Logical Fallacies	16/20
World Religions	18/20
Total	158/200 (79%)

JANG_2S Profile

JANG_2S is an aggressive 2-bit mixed-precision profile that makes the 122B parameter model runnable on a single Mac. Critical layers (attention, routing, embeddings) are protected at higher precision while expert MLP weights are compressed to 2-bit.

Usage

# Requires Osaurus (https://osaurus.ai)
osaurus serve OsaurusAI/Qwen3.5-122B-A10B-JANG_2S

Requirements

• Apple Silicon Mac with 48+ GB unified memory
• MLX framework with Qwen 3.5 MoE support

---

Quantized by Osaurus AI using JANG