MERaLiON-2-3B-RotorQuant-MLX-2bit

MLX 2-bit RotorQuant quantization of aisingapore/MERaLiON-AudioLLM-Whisper-SEA-LION-V3-3B for Apple Silicon inference.

RotorQuant applies rotation-based quantization that decorrelates weight matrices before quantization, distributing outlier magnitudes more evenly across channels for improved accuracy at low bit-widths.

Model Specifications

Property	Value
Base Model	aisingapore/MERaLiON-AudioLLM-Whisper-SEA-LION-V3-3B
Parameters	~3B
Architecture	Whisper-large-v3 encoder + Gemma-2-2B-IT decoder
Quantization	RotorQuant 2-bit (MLX)
Disk Size	~1 GB
Peak RAM	~1.5 GB
License	Apache 2.0
Task	Automatic Speech Recognition / Speech-to-Text

Quickstart

Installation

pip install mlx-lm mlx-whisper

Inference

from mlx_lm import load, generate
from mlx_lm.cache import IsoQuantCache

model, tokenizer = load("majentik/MERaLiON-2-3B-RotorQuant-MLX-2bit")

# Create IsoQuantCache for RotorQuant models
cache = IsoQuantCache(model)

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Transcribe the following audio."}],
    tokenize=False,
    add_generation_prompt=True,
)

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=512,
    cache=cache,
)
print(response)

Quantization Details

RotorQuant is a rotation-based quantization strategy that:

• Applies learned rotation matrices to decorrelate weight channels before quantization
• Reduces the impact of outlier weights that typically degrade quantized model quality
• Provides more uniform weight distributions, leading to better accuracy retention
• Pairs with IsoQuantCache for consistent KV-cache quantization during inference

This 2-bit variant offers the smallest possible footprint for the 3B model. RotorQuant's rotation-based approach is especially valuable at 2-bit, where outlier sensitivity causes the most quality degradation in naive quantization schemes. This makes it the preferred 2-bit option when accuracy matters.

Supported Languages

MERaLiON-2 supports speech recognition in Southeast Asian languages including English, Mandarin Chinese, Malay, Tamil, and Indonesian.

Memory Estimates

Device	Feasibility
MacBook Air M1 (8 GB)	Comfortable
iPad Pro M1/M2	Comfortable
iPad Air M1	Feasible
iPhone 15 Pro (8 GB)	Feasible

See Also

• majentik/MERaLiON-2-3B-TurboQuant-MLX-2bit -- TurboQuant 2-bit variant
• majentik/MERaLiON-2-3B-RotorQuant-MLX-4bit -- RotorQuant 4-bit (higher quality)
• majentik/MERaLiON-2-3B-RotorQuant-MLX-8bit -- RotorQuant 8-bit (highest quality)
• majentik/MERaLiON-2-10B-RotorQuant-MLX-2bit -- 10B RotorQuant 2-bit (larger model)
• aisingapore/MERaLiON-AudioLLM-Whisper-SEA-LION-V3-3B -- Original base model

Quant trade-off (MLX lane)

Bits	Approx size	Use case	Recommendation
2-bit	~799 MB	Aggressive quantization	Very low-RAM Macs
3-bit	~1.1 GB	Lossy but small	Low-RAM Macs
4-bit	~1.3 GB	Balanced default	Recommended for most Macs
5-bit	~1.5 GB	Higher fidelity	Quality-sensitive
6-bit	~1.8 GB	Approaching FP16 quality	High-fidelity
8-bit	~2.3 GB	Near-lossless reference	Fidelity-critical work

(Current variant — 2bit — is bolded.)

Variants in this family

(Showing 8 sibling variants under majentik/meralion2-3b-*. The current variant — RotorQuant-MLX-2bit — is bolded.)

Variant	Runtime	Approx size	Use case
RotorQuant	runtime modifier	n/a	KV-cache root (weight-agnostic)
RotorQuant-MLX-2bit	mlx-lm	~983 MB	Apple Silicon, smallest
RotorQuant-MLX-4bit	mlx-lm	~1.9 GB	Apple Silicon balanced
RotorQuant-MLX-8bit	mlx-lm	~3.5 GB	Apple Silicon reference
TurboQuant	runtime modifier	n/a	KV-cache root (weight-agnostic)
TurboQuant-MLX-2bit	mlx-lm	~983 MB	Apple Silicon, smallest
TurboQuant-MLX-4bit	mlx-lm	~1.9 GB	Apple Silicon balanced
TurboQuant-MLX-8bit	mlx-lm	~3.5 GB	Apple Silicon reference