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-2-10B-TurboQuant-MLX-2bit

MLX 2-bit TurboQuant quantization of aisingapore/MERaLiON-AudioLLM-Whisper-SEA-LION-V3-10B for Apple Silicon inference.

TurboQuant applies mixed-precision quantization that preserves critical attention layers at higher precision while aggressively quantizing less sensitive feed-forward layers, optimizing for speed without sacrificing quality.

Model Specifications

Property	Value
Base Model	aisingapore/MERaLiON-AudioLLM-Whisper-SEA-LION-V3-10B
Parameters	~10B
Architecture	Whisper encoder + Gemma-2-9B-IT decoder
Quantization	TurboQuant 2-bit (MLX)
Disk Size	~3 GB
Peak RAM	~4 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 TurboQuantCache

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

# Create TurboQuant-optimized KV cache
cache = TurboQuantCache(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

TurboQuant is a mixed-precision quantization strategy that:

• Retains attention projection layers at higher precision
• Quantizes MLP/feed-forward layers more aggressively where precision loss is tolerable
• Optimizes KV-cache memory layout for faster autoregressive decoding on Apple Silicon

This 2-bit variant offers the smallest memory footprint, making it possible to run a 10B-parameter speech model on devices with as little as 8 GB of unified memory. Expect some quality degradation compared to 4-bit and 8-bit variants.

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)	Feasible
MacBook Pro M1/M2 (16 GB)	Comfortable
MacBook Pro M2/M3 (32 GB)	Ideal
Mac Studio M2 Ultra (64 GB+)	Overkill for this variant

See Also

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

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/meralion2-10b-*. The current variant — TurboQuant-MLX-2bit — is bolded.)

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

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.