MERaLiON-3-10B-RotorQuant-MLX-8bit / README.md

majentik

chore(card): enrich YAML frontmatter (pipeline_tag, language, library_name, inference)

ef6b66b verified 1 day ago

preview code

raw

history blame contribute delete

3.67 kB

metadata

base_model: MERaLiON/MERaLiON-3-10B-preview
library_name: mlx
tags:
  - rotorquant
  - kv-cache-quantization
  - meralion
  - speech-to-text
  - multimodal
  - audio
  - quantized
  - mlx
  - 8bit
  - apple-silicon
license: other
pipeline_tag: automatic-speech-recognition
language:
  - en

MERaLiON-3-10B-RotorQuant-MLX-8bit

8-bit weight-quantized MLX version of MERaLiON/MERaLiON-3-10B-preview with RotorQuant KV-cache quantization. Optimized for Apple Silicon inference via the MLX framework.

MERaLiON-3-10B is a multimodal audio-language model built on a Gemma-2 decoder backbone, designed for speech-to-text and audio understanding tasks.

Approximate model size: ~10 GB

Model Specifications

Property	Value
Base Model	MERaLiON/MERaLiON-3-10B-preview
Parameters	~10 billion
Architecture	Multimodal audio-language (Gemma-2 decoder backbone)
Modality	Audio + text input, text output
License	See base model
Weight Quantization	8-bit (~10 GB)
KV-Cache Quantization	RotorQuant
Framework	MLX (Apple Silicon)

Quickstart

from mlx_lm import load, generate

model, tokenizer = load("majentik/MERaLiON-3-10B-RotorQuant-MLX-8bit")

prompt = "Transcribe the following audio:"
response = generate(model, tokenizer, prompt=prompt, max_tokens=512)
print(response)

What is RotorQuant?

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

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

Combined with MLX 8-bit weight quantization, this dual compression approach provides excellent throughput for audio processing workloads.

KV-Cache Quantization Comparison

Method	Prefill Speed	Decode Speed	Memory Savings	Reference
TurboQuant	Baseline	Baseline	High	arXiv: 2504.19874
RotorQuant	5.3x faster	28% faster	High	GitHub

Memory Estimates (MERaLiON-3-10B)

Precision	Approximate Size	MLX Variant
FP16 (original)	~20 GB	--
8-bit quantized	~10 GB	This model
4-bit quantized	~5 GB	RotorQuant-MLX-4bit
2-bit quantized	~3 GB	RotorQuant-MLX-2bit

Hardware Requirements

This model requires approximately 10 GB of unified memory. Recommended hardware:

Apple M1 Pro (16 GB+)
Apple M2/M3/M4 (16 GB+)
Any Apple Silicon Mac with 16 GB+ RAM

majentik
/

MERaLiON-3-10B-RotorQuant-MLX-8bit