Mistral-Small-4-119B-RotorQuant-MLX-4bit / README.md

Add MLX 4-bit quantized Mistral Small 4 with KV cache compression

9968358 verified 3 days ago

3.57 kB

base_model: mistralai/Mistral-Small-4-119B-2603
library_name: mlx
license: apache-2.0
tags:
  - rotorquant
  - kv-cache-quantization
  - mistral
  - moe
  - sparse-moe
  - multimodal
  - quantized
  - mlx
  - 4-bit
  - apple-silicon
  - 256k-context
  - thinking
pipeline_tag: text-generation

Mistral-Small-4-119B-RotorQuant-MLX-4bit

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

This repository provides a 4-bit weight-quantized MLX conversion of mistralai/Mistral-Small-4-119B-2603 with RotorQuant KV cache quantization support. Designed for efficient inference on Apple Silicon Macs.

Overview

This model applies two complementary compression techniques:

4-bit weight quantization (MLX) -- reduces model weights from ~238 GB to ~60 GB
RotorQuant KV cache quantization -- reduces KV cache from ~32 GB to ~6.5 GB at 256K context

Together, these make it feasible to run a 119B-parameter MoE model on high-memory Apple Silicon machines with excellent throughput.

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	4-bit (MLX)
KV Cache Quantization	RotorQuant 3-bit

Memory Estimates

Configuration	Weights	KV Cache (256K)	Total
FP16 baseline	~238 GB	~32 GB	~270 GB
This model (4-bit MLX + RotorQuant)	~60 GB	~6.5 GB	~66.5 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.

Quickstart

from mlx_lm import load, generate

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

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)

What is RotorQuant?

RotorQuant is a rotation-based KV cache quantization method that applies learned rotations before quantizing the key-value cache. Key results on the base model:

5.3x faster prefill compared to unquantized baseline
28% faster decode throughput
Perplexity: 6.91 vs 7.07 for unquantized (lower is better)

Because it targets the KV cache rather than weights, it stacks with weight quantization for compounding memory savings.

majentik
/

Mistral-Small-4-119B-RotorQuant-MLX-4bit