[!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.

MERaLiON-3-10B-RotorQuant-MLX-2bit

2-bit weight-quantized MLX version of MERaLiON/MERaLiON-3-10B-preview with the legacy RotorQuant KV-cache fork (superseded by upstream llama.cpp KV options). 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: ~3 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	2-bit (~3 GB)
KV-Cache Quantization	RotorQuant
Framework	MLX (Apple Silicon)

Quickstart

python

from mlx_lm import load, generate

model, tokenizer = load("majentik/MERaLiON-3-10B-RotorQuant-MLX-2bit")

prompt = "Transcribe the following audio:"
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	Baseline	Baseline	High	arXiv: 2504.19874

Memory Estimates (MERaLiON-3-10B)

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

Hardware Requirements

This model requires approximately 3 GB of unified memory. Recommended hardware:

Apple M1 (8 GB+)
Any Apple Silicon Mac

Quant trade-off (MLX lane)

Bits	Approx size	Use case	Recommendation
2-bit	~2.6 GB	Aggressive quantization	Very low-RAM Macs
3-bit	~3.6 GB	Lossy but small	Low-RAM Macs
4-bit	~4.2 GB	Balanced default	Recommended for most Macs
5-bit	~5.0 GB	Higher fidelity	Quality-sensitive
6-bit	~6.0 GB	Approaching FP16 quality	High-fidelity
8-bit	~7.6 GB	Near-lossless reference	Fidelity-critical work

(Current variant — 2bit — is bolded.)

Variants in this family

(Showing 8 sibling variants under majentik/meralion3-10b-*. The current variant — RotorQuant-MLX-2bit — is bolded.)

Variant	Runtime	Approx size	Use case
RotorQuant-MLX-2bit	mlx-lm	~3.2 GB	Apple Silicon, smallest

majentik/MERaLiON-3-10B-RotorQuant-MLX-2bit