majentik/Qwen3.5-27B-TurboQuant-MLX-2bit

[!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-TurboQuant-MLX-2bit

MLX 2-bit weight quantization + TurboQuant 2-bit KV cache compression for Qwen/Qwen3.5-27B.

Dual compression for Apple Silicon: both the model weights and the KV cache are quantized to 2-bit, enabling long-context inference on memory-constrained Macs.

Overview

Qwen3.5-27B is a 27B-parameter hybrid transformer with 262K native context and built-in thinking mode (the model generates internal reasoning tokens before answering). Thinking mode makes KV cache compression especially valuable, since the reasoning chain can consume substantial cache memory.

This variant applies two layers of compression:

1. 1.MLX 2-bit weight quantization — reduces the 27B model from ~54 GB (BF16) to approximately ~8 GB, making it loadable on Apple Silicon devices with limited unified memory.
2. 2.TurboQuant 2-bit KV cache — compresses the key-value cache by approximately 8x compared to FP16, enabling long-context inference without running out of memory.

At 2-bit precision, both weight and cache quantization are aggressive — expect some quality degradation compared to 4-bit variants, but this combination enables running a 27B thinking model with long context on hardware that would otherwise be unable to fit it.

Specifications

Memory Estimates

Quickstart

from mlx_lm import load, generate
from turboquant import TurboQuantCache

model_id = "majentik/Qwen3.5-27B-TurboQuant-MLX-2bit"

model, tokenizer = load(model_id)

# Apply 2-bit KV cache compression
cache = TurboQuantCache(bits=2)

prompt = "Explain the Riemann hypothesis in simple terms."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(
    model,
    tokenizer,
    prompt=text,
    max_tokens=2048,
    kv_cache=cache,
)
print(response)

Quality Notes

• —2-bit weights + 2-bit KV cache is the most aggressive quantization combination. Use this when memory is the primary constraint and some quality loss is acceptable.
• —For higher quality on Apple Silicon, consider 4-bit weight variants with 4-bit KV cache.
• —Thinking mode reasoning quality may be more sensitive to quantization since the model relies on both weight precision and cached reasoning tokens for its final answer.
• —Best suited for: prototyping, development, long-context exploration, and scenarios where running the model at all matters more than peak quality.

References

• —TurboQuant: Efficient KV Cache Quantization
• —MLX — Apple's machine learning framework
• —mlx-lm — LLM inference with MLX
• —Qwen3.5-27B base model

See Also

Quant trade-off (MLX lane)

(Current variant — 2bit — is bolded.)

Variants in this family

(Showing 16 sibling variants under majentik/qwen3.5-27b-*. The current variant — TurboQuant-MLX-2bit — is bolded.)

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.