majentik/Nemotron-3-Nano-4B-TurboQuant-MLX-2bit

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.

Nemotron-3-Nano-4B - TurboQuant MLX 2-bit

2-bit weight-quantized MLX version of nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 with the legacy TurboQuant KV-cache fork (superseded by upstream llama.cpp KV options). Optimized for Apple Silicon inference via the MLX framework. Maximum compression for running on memory-constrained devices. The dense hybrid Mamba-2 + Attention architecture supports up to 262K context length.

Approximate model size: ~1.2 GB

Model Specifications

Property	Value
Base Model	nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16
Parameters	4 billion (dense)
Architecture	Hybrid Mamba-2 + Attention (dense)
Context Length	262,144 tokens (262K)
License	NVIDIA Open Model License (commercial use OK)
Weight Quantization	2-bit (~1.2 GB)
KV-Cache Quantization	TurboQuant
Framework	MLX (Apple Silicon)

Quickstart

from mlx_lm import load, generate
from turboquant import TurboQuantCache

model, tokenizer = load("majentik/Nemotron-3-Nano-4B-TurboQuant-MLX-2bit")

prompt = "Explain the theory of relativity."
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	1x (baseline)	1x (baseline)	High	arXiv: 2504.19874

Memory Estimates (Nemotron-3-Nano-4B)

Precision	Approximate Size	MLX Variant
BF16 (original)	~8 GB	--
8-bit quantized	~4 GB	TurboQuant-MLX-8bit
4-bit quantized	~2.3 GB	TurboQuant-MLX-4bit
2-bit quantized	~1.2 GB	This model

Hardware Requirements

This model requires approximately 1.2 GB of unified memory. Recommended hardware:

• Apple M1 (8 GB+)
• Apple M2 (8 GB+)
• Apple M3 (8 GB+)
• Apple M4 (8 GB+)
• Any Apple Silicon Mac with 8 GB+ unified memory

See Also

• nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 -- Base model
• majentik/Nemotron-3-Nano-4B-TurboQuant-MLX-8bit -- MLX 8-bit variant
• majentik/Nemotron-3-Nano-4B-TurboQuant-MLX-4bit -- MLX 4-bit variant
• majentik/Nemotron-3-Nano-4B-RotorQuant-MLX-2bit -- RotorQuant MLX 2-bit variant
• TurboQuant Paper (arXiv: 2504.19874)
• MLX Framework

Quant trade-off (MLX lane)

Bits	Approx size	Use case	Recommendation
2-bit	~1.0 GB	Aggressive quantization	Very low-RAM Macs
3-bit	~1.4 GB	Lossy but small	Low-RAM Macs
4-bit	~1.7 GB	Balanced default	Recommended for most Macs
5-bit	~2.0 GB	Higher fidelity	Quality-sensitive
6-bit	~2.4 GB	Approaching FP16 quality	High-fidelity
8-bit	~3.0 GB	Near-lossless reference	Fidelity-critical work

(Current variant — 2bit — is bolded.)

Variants in this family

(Showing 13 sibling variants under majentik/nemotron3-nano-4b-*. The current variant — TurboQuant-MLX-2bit — is bolded.)

Variant	Runtime	Approx size	Use case
TurboQuant-MLX-2bit	mlx-lm	~1.3 GB	Apple Silicon, smallest