[!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.

GPT-OSS-120B - TurboQuant MLX 4-bit

4-bit weight-quantized MLX version of openai/gpt-oss-120b with the legacy TurboQuant KV-cache fork (superseded by upstream llama.cpp KV options). Optimized for Apple Silicon inference via the MLX framework. A good balance between model quality and memory efficiency, enabling GPT-OSS-120B to fit on high-memory Mac Studios. GPT-OSS-120B is OpenAI's flagship open-weights Mixture-of-Experts model (Apache 2.0), approaching o4-mini quality for reasoning tasks.

Approximate model size: ~65 GB

Model Specifications

Property	Value
Base Model	openai/gpt-oss-120b
Parameters	120 billion (MoE)
Architecture	Mixture-of-Experts (MoE) Transformer
License	Apache 2.0 (commercial use OK)
Weight Quantization	4-bit (~65 GB)
KV-Cache Quantization	TurboQuant
Framework	MLX (Apple Silicon)

Quickstart

python

from mlx_lm import load, generate
from turboquant import TurboQuantCache

model, tokenizer = load("majentik/gpt-oss-120b-TurboQuant-MLX-4bit")

prompt = "Explain the theory of relativity."
response = generate(model, tokenizer, prompt=prompt, max_tokens=512)
print(response)

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).

KV-Cache Quantization Comparison

Method	Prefill Speed	Decode Speed	Memory Savings	Reference
TurboQuant	1x (baseline)	1x (baseline)	High	arXiv: 2504.19874

Memory Estimates (GPT-OSS-120B)

Precision	Approximate Size	MLX Variant
BF16 (original)	~240 GB	--
4-bit quantized	~65 GB	This model

Hardware Requirements

This model requires approximately 65 GB of unified memory. Recommended hardware:

Apple M2 Max (96 GB)
Apple M2 Ultra (96 GB or 192 GB)
Apple M3 Max (96 GB+) / M3 Ultra (192 GB+)
Apple M4 Max / M4 Ultra (96 GB+)

Quant trade-off (MLX lane)

Bits	Approx size	Use case	Recommendation
2-bit	~31 GB	Aggressive quantization	Very low-RAM Macs
3-bit	~43 GB	Lossy but small	Low-RAM Macs
4-bit	~50 GB	Balanced default	Recommended for most Macs
5-bit	~60 GB	Higher fidelity	Quality-sensitive
6-bit	~72 GB	Approaching FP16 quality	High-fidelity
8-bit	~91 GB	Near-lossless reference	Fidelity-critical work

(Current variant — 4bit — is bolded.)

Variants in this family

(Showing 14 sibling variants under majentik/gpt-oss-120b-*. The current variant — TurboQuant-MLX-4bit — is bolded.)

Variant	Runtime	Approx size	Use case
TurboQuant-MLX-4bit	mlx-lm	~74 GB	Apple Silicon balanced

majentik/gpt-oss-120b-TurboQuant-MLX-4bit