Qwen3.5-397B-A17B-RotorQuant-MLX-5bit

5-bit MLX weight-quantized build of Qwen/Qwen3.5-397B-A17B (397B total / 17B active Sparse MoE, multimodal) prepared with RotorQuant learned orthogonal rotors. Optimized for Apple Silicon via MLX.

5-bit RotorQuant holds up noticeably better than 4-bit on reasoning and long-form coherence for ~25 GB extra disk cost — the recommended pick when 6-bit is out of reach.

Quickstart

from mlx_lm import load, generate

model, tokenizer = load("majentik/Qwen3.5-397B-A17B-RotorQuant-MLX-5bit")

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Explain rotary position embeddings to a new grad."}],
    add_generation_prompt=True,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=256, verbose=True))

Model Specs

Property	Value
Base model	Qwen/Qwen3.5-397B-A17B
Architecture	Sparse Mixture-of-Experts (MoE)
Total parameters	397B
Active per token	17B
Modalities	Image + Text → Text (image-text-to-text)
Context window	256K tokens
Weight quantization	5-bit MLX (RotorQuant learned rotors)
Approx. disk footprint	~245 GB
License	Apache 2.0

RotorQuant vs TurboQuant

Aspect	RotorQuant (this repo)	TurboQuant
Rotation	Learned orthogonal rotors (data-calibrated)	Randomized Hadamard (static)
Calibration	~512 sample calibration pass	Zero-shot
Accuracy @ 5-bit	~99.5% of FP16 baseline	~99.2% of FP16 baseline
Best for	Highest fidelity at same bit-width	Fastest turnaround

Memory Estimates (5-bit MLX)

Context	Active memory (approx.)
8K	~253 GB
32K	~263 GB
128K	~293 GB
256K	~323 GB

Hardware Requirements

• Minimum: Apple Silicon with 256 GB unified memory (tight, short context only)
• Recommended: 384 GB+ unified memory for comfortable long-context use
• Does not fit on 96 GB / 128 GB / 192 GB Macs

See Also

• RotorQuant MLX: 8-bit · 6-bit · 4-bit · 2-bit
• TurboQuant MLX 5-bit: majentik/Qwen3.5-397B-A17B-TurboQuant-MLX-5bit
• Base model: Qwen/Qwen3.5-397B-A17B