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

[!TIP] 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.

<!-- kv-upstream-note -->

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

About the RotorQuant / TurboQuant labels

RotorQuant and TurboQuant are this project's release labels, not distinct quantization algorithms — for any given tier, both brand repos carry byte-identical weights produced with the standard MLX / llama.cpp quantizers. No brand-specific speedup is claimed or measured. The KV-cache fork these labels originally referred to is legacy; for KV-cache memory savings use the upstream options described above (-ctk/-ctv q8_0, OLLAMA_KV_CACHE_TYPE).

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

See Also

• Base model
• RotorQuant GitHub

Quant trade-off (MLX lane)

(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.)