[!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_0withOLLAMA_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-30B-A3B - RotorQuant MLX 4-bit
4-bit weight-quantized MLX version of nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 with the legacy RotorQuant KV-cache fork (superseded by upstream llama.cpp KV options). Optimized for Apple Silicon inference via the MLX framework. A good balance between model quality and memory efficiency. Only 3.2B parameters are active per token despite 30.7B total, making this model significantly more efficient at inference time than its parameter count suggests. The hybrid Mamba-2 + Transformer MoE architecture supports up to 1M context length.
Approximate model size: ~17 GB
Model Specifications
| Property | Value |
|---|---|
| Base Model | nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 |
| Parameters | 30.7 billion total (3.2 billion active per token) |
| Architecture | Hybrid Mamba-2 + Transformer MoE (3.2B active per token) |
| Context Length | 1,048,576 tokens (1M) |
| License | NVIDIA Open Model License (commercial use OK) |
| Weight Quantization | 4-bit (~17 GB) |
| KV-Cache Quantization | RotorQuant |
| Framework | MLX (Apple Silicon) |
Quickstart
from mlx_lm import load, generate
from rotorquant import IsoQuantCache
model, tokenizer = load("majentik/Nemotron-3-Nano-30B-A3B-RotorQuant-MLX-4bit")
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-30B-A3B)
| Precision | Approximate Size | MLX Variant |
|---|---|---|
| BF16 (original) | ~60 GB | -- |
| 8-bit quantized | ~30 GB | RotorQuant-MLX-8bit |
| 4-bit quantized | ~17 GB | This model |
Hardware Requirements
This model requires approximately 17 GB of unified memory. Recommended hardware:
- Apple M2 Pro (24 GB+)
- Apple M3 Pro (24 GB+)
- Apple M4 Pro (24 GB+)
- Any Apple Silicon Mac with 24 GB+ unified memory
See Also
- nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 -- Base model
- majentik/Nemotron-3-Nano-30B-A3B-RotorQuant-MLX-8bit -- MLX 8-bit variant
- RotorQuant GitHub
- MLX Framework