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.

Mistral-Small-4-119B-TurboQuant-MLX-2bit

Dual compression: 2-bit MLX weight quantization + TurboQuant KV cache quantization for Mistral Small 4 on Apple Silicon.

This repository provides a 2-bit weight-quantized MLX conversion of mistralai/Mistral-Small-4-119B-2603 with TurboQuant KV cache quantization support. Aggressive compression for running on consumer Apple Silicon hardware.

Overview

This model applies two complementary compression techniques:

1. 2-bit weight quantization (MLX) -- reduces model weights from ~238 GB to ~30 GB
2. TurboQuant KV cache quantization -- reduces KV cache from ~32 GB to ~8 GB at 256K context

This enables running a 119B-parameter MoE model on Apple Silicon Macs with 64 GB+ unified memory.

Model Specs

Property	Value
Base Model	Mistral Small 4 (March 2026)
Total Parameters	119B
Active Parameters	6.5B per token (Sparse MoE)
Architecture	Sparse MoE -- 128 experts, 4 active per token
Context Length	256K tokens
Modality	Text + Images (multimodal)
Capabilities	Thinking / reasoning, tool use, multilingual
License	Apache 2.0
Weight Quantization	2-bit (MLX)
KV Cache Quantization	TurboQuant 4-bit

Memory Estimates

Configuration	Weights	KV Cache (256K)	Total
FP16 baseline	~238 GB	~32 GB	~270 GB
This model (2-bit MLX + TurboQuant)	~30 GB	~8 GB	~38 GB

Note: This is a Sparse MoE model -- only 6.5B parameters are active per token, so inference is fast despite the 119B total parameter count. The 2-bit quantization trades some quality for significantly reduced memory. Expect modest degradation on complex reasoning tasks compared to 4-bit.

Quickstart

from mlx_lm import load, generate

model, tokenizer = load("majentik/Mistral-Small-4-119B-TurboQuant-MLX-2bit")

prompt = "Explain sparse mixture-of-experts architectures."
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=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).

See Also

• mistralai/Mistral-Small-4-119B-2603 -- Base model
• majentik/Mistral-Small-4-119B-TurboQuant-MLX-4bit -- 4-bit MLX variant
• TurboQuant Paper (arXiv: 2504.19874)