KV-cache quantization without any fork (recommended, 2026): upstream llama.cpp/Ollama now cover this natively — use -ctk q8_0 -ctv q8_0 (half KV memory, negligible quality loss: perplexity +0.002–0.05) or -ctk q4_0 -ctv q4_0 (quarter memory, ≈7.6% perplexity increase). In Ollama: OLLAMA_KV_CACHE_TYPE=q8_0 with OLLAMA_FLASH_ATTENTION=1. Keep K and V types symmetric to stay on the fast fused Flash-Attention path. Since April 2026, mainline llama.cpp also applies Hadamard rotation to KV activations (PR #21038), which greatly improves low-bit KV quality (opt-out: LLAMA_ATTN_ROT_DISABLE=1).

The RotorQuant/TurboQuant fork flow below is experimental/legacy: the TurboQuant llama.cpp PR was closed without merging (June 2026) and the fork is unmaintained relative to mainline. It is NOT required to use this model.

Qwen3.5-27B-RotorQuant-MLX-8bit

MLX 8-bit weight-quantized variant of Qwen/Qwen3.5-27B with RotorQuant KV cache compression for efficient inference on Apple Silicon.

Overview

This model combines two complementary compression techniques:

• MLX 8-bit weight quantization — reduces model size from ~54GB to ~27GB
• RotorQuant KV cache compression — compresses key-value caches during inference using Clifford algebra block-diagonal rotations, enabling longer contexts with less VRAM

Quickstart

from mlx_lm import load, generate
from turboquant import IsoQuantCache

model, tokenizer = load("majentik/Qwen3.5-27B-RotorQuant-MLX-8bit")

# Standard generation
prompt = "Explain the theory of relativity"
response = generate(model, tokenizer, prompt=prompt, max_tokens=2048)
print(response)

Specifications

Property	Value
Base Model	Qwen/Qwen3.5-27B
Parameters	27B
Weight Quantization	MLX 8-bit affine
KV Cache Method	RotorQuant (Clifford algebra block-diagonal rotations)
Model Size	~27 GB
Context Length	262K (native), 1M+ (extended)
Platform	Apple Silicon (M1/M2/M3/M4/M5)

What is RotorQuant?

RotorQuant uses Clifford algebra block-diagonal rotations for KV cache quantization, achieving superior efficiency compared to vector-quantization-based approaches like TurboQuant:

Metric	RotorQuant	TurboQuant
Prefill speed	5.3x faster	baseline
Decode speed	28% faster	baseline
Quantizer parameters	44x fewer	baseline
Perplexity	6.91	7.07

RotorQuant compresses the KV cache to ~3-bit effective precision while maintaining lower perplexity than TurboQuant's 4-bit approach, thanks to the mathematical efficiency of geometric algebra rotations.

Thinking Mode

Qwen3.5-27B generates extended reasoning before responses by default. The combination of weight quantization and KV cache compression is especially valuable here — thinking tokens consume significant memory that is reduced by both techniques working together.

Memory Estimate

Configuration	Model Weights	KV Cache (128K ctx)	Total
FP16 (baseline)	~54 GB	~13 GB	~67 GB
MLX 8-bit + RotorQuant	~27 GB	~1.3 GB	~28.3 GB

See Also

• Base model
• KV-cache-only variant
• RotorQuant GitHub

Quant trade-off (MLX lane)

Bits	Approx size	Use case	Recommendation
2-bit	~7.3 GB	Aggressive quantization	Very low-RAM Macs
3-bit	~10 GB	Lossy but small	Low-RAM Macs
4-bit	~12 GB	Balanced default	Recommended for most Macs
5-bit	~14 GB	Higher fidelity	Quality-sensitive
6-bit	~17 GB	Approaching FP16 quality	High-fidelity
8-bit	~21 GB	Near-lossless reference	Fidelity-critical work

(Current variant — 8bit — is bolded.)

Variants in this family

(Showing 16 sibling variants under majentik/qwen3.5-27b-*. The current variant — RotorQuant-MLX-8bit — is bolded.)

Variant	Runtime	Approx size	Use case
RotorQuant	runtime modifier	n/a	KV-cache root (weight-agnostic)
RotorQuant-2bit	transformers	n/a	Standalone 2-bit weights
RotorQuant-GGUF-IQ4_XS	llama.cpp	~23 GB	Lossy 4-bit, low-RAM CPU/edge
RotorQuant-GGUF-Q2_K	llama.cpp	~16 GB	Lossy, low-RAM CPU/edge
RotorQuant-GGUF-Q3_K_M	llama.cpp	~21 GB	Smaller 3-bit, CPU-friendly
RotorQuant-GGUF-Q4_K_M	llama.cpp	~30 GB	Balanced default
RotorQuant-GGUF-Q5_K_M	llama.cpp	~36 GB	Higher fidelity, more RAM
RotorQuant-GGUF-Q8_0	llama.cpp	~57 GB	Near-lossless reference
RotorQuant-MLX-2bit	mlx-lm	~8.6 GB	Apple Silicon, smallest
RotorQuant-MLX-4bit	mlx-lm	~17 GB	Apple Silicon balanced
RotorQuant-MLX-8bit	mlx-lm	~32 GB	Apple Silicon reference
TurboQuant	runtime modifier	n/a	KV-cache root (weight-agnostic)
TurboQuant-2bit	transformers	n/a	Standalone 2-bit weights
TurboQuant-MLX-2bit	mlx-lm	~8.6 GB	Apple Silicon, smallest
TurboQuant-MLX-4bit	mlx-lm	~17 GB	Apple Silicon balanced
TurboQuant-MLX-8bit	mlx-lm	~32 GB	Apple Silicon reference