Tinker-Stack/Nemotron-3-Nano-30B-A3B-IQ4_XS-GGUF

NVIDIA Nemotron-3-Nano-30B-A3B — IQ4_XS GGUF (RTX 2080 Ti Optimized)

Production-validated IQ4_XS quantization of NVIDIA Nemotron-3-Nano-30B-A3B, benchmarked and tuned for 11GB VRAM GPUs (RTX 2080 Ti / RTX 3060 / etc).

GGUF sourced from bartowski/nvidia_Nemotron-3-Nano-30B-A3B-GGUF. This repo adds real-world benchmarks, production configs, and deployment guides for constrained-VRAM setups.

Why This Repo?

bartowski provides the quants — we provide the deployment playbook for running this model on a single consumer GPU with real benchmark data, not projections.

• ✅ Validated tool calling (function calls parse correctly)
• ✅ Validated reasoning (chain-of-thought with <think> tags)
• ✅ Production Ollama Modelfile included
• ✅ llama-server launch configs with measured VRAM headroom
• ✅ Tested across 168 messages in a multi-agent war room

Model Details

Property	Value
Base Model	NVIDIA Nemotron-3-Nano-30B-A3B
Architecture	MoE Hybrid: 23 Mamba-2 + 6 Attention + 128 experts + 1 shared
Total Parameters	31.6B
Active Parameters	~3.2–3.6B (Mixture of Experts)
Quantization	IQ4_XS (imatrix)
File Size	16.8 GB
Format	GGUF
Prompt Format	ChatML (<|im_start|>)
Context Window	Up to 5120 tested stable on 11GB (hardware dependent)
License	NVIDIA Open Model License

Benchmarks (RTX 2080 Ti, 11GB VRAM)

All benchmarks on: Ryzen 7 3800X (8c/16t) · 32GB DDR4-3600 · RTX 2080 Ti 11GB · Ubuntu 24.04

Throughput

Backend	GPU Layers	KV Cache	Context	Flash Attn	tok/s (single)	tok/s (sustained)
Ollama (Q4_K_M)	28/52 (est.)	default	4096	❌	11.04	—
Ollama (IQ4_XS)	28/52	default	4096	❌	14.49	—
llama-server (IQ4_XS)	28	q8_0	4096	❌	22.85	19.45–19.69
llama-server (IQ4_XS)	28	q8_0	4096	✅	26.25	26.21–26.70

VRAM Usage

Config	VRAM Used	Free	Headroom
Ollama Q4_K_M	10,449 MiB	815 MiB	⚠️ Tight
Ollama IQ4_XS	10,079 MiB	1,185 MiB	✅ Good
llama-server ctx 2048 (×1 slot)	10,631 MiB	633 MiB	✅ Safe
llama-server ctx 3072 (×1 slot)	10,627 MiB	637 MiB	✅ Safe
llama-server ctx 4096 (×1 slot)	10,639 MiB	625 MiB	✅ Production
llama-server ctx 5120 (×1 slot)	10,641 MiB	623 MiB	✅ Safe
llama-server ctx 6144 (×1 slot)	10,947 MiB	317 MiB	⚠️ Tight
llama-server ctx 8192 (×1 slot)	11,085 MiB	179 MiB	❌ OOM risk
llama-server ctx 4096 (×4 slots)	~10,953 MiB	311 MiB	⚠️ Dangerous

Why does context barely affect VRAM? Nemotron's hybrid Mamba-2 architecture uses recurrent state (fixed size) for 46 of 52 layers. Only 6 attention layers need KV cache, so context scaling is nearly flat until ~6K where compute buffers jump.

IQ4_XS vs Q4_K_M (Same Model, Ollama)

Metric	Q4_K_M	IQ4_XS	Delta
Throughput	11.04 tok/s	14.49 tok/s	+31%
VRAM	10,449 MiB	10,079 MiB	−370 MiB
Cold Start	42.7s	23.7s	−44%
Disk Size	24 GB	16.8 GB	−6 GB
Tool Calling	✅	✅	—
Reasoning	✅	✅	—

Quality Validation

• Math reasoning: "What's 15% tip on $28.50?" → Correct ($4.28 → $4.275, rounds properly)
• Tool calling: OpenAI-format function calls parse correctly
• Multi-turn: Sustained coherence over 168-message war room session

Recommended Configurations

Ollama (Easy, Recommended for Most Users)

Create a Modelfile:

FROM ./nvidia_Nemotron-3-Nano-30B-A3B-IQ4_XS.gguf

PARAMETER temperature 0.3
PARAMETER top_k 40
PARAMETER top_p 0.85
PARAMETER num_ctx 4096
PARAMETER num_predict 512
PARAMETER repeat_penalty 1.1

ollama create nemotron-prod -f Modelfile
ollama run nemotron-prod

llama-server (Maximum Performance, +84% over Ollama)

# Production config (recommended for 11GB VRAM)
llama-server \
  --model nvidia_Nemotron-3-Nano-30B-A3B-IQ4_XS.gguf \
  --n-gpu-layers 28 \
  --cache-type-k q8_0 \
  --ctx-size 4096 \
  --parallel 1 \
  --flash-attn on \
  --port 8081

Flash Attention: Build llama.cpp with GGML_CUDA_FA=ON and use --flash-attn on (not auto — auto may fail to enable on split GPU/CPU models). FA gives +35% sustained throughput on this model.

28 GPU layers is the stable ceiling for 11GB VRAM. 29 loads but OOMs during generation. 30+ fails at load.

Context window scaling (measured on RTX 2080 Ti, single slot):

• --ctx-size 2048 → 633 MiB free ✅
• --ctx-size 3072 → 637 MiB free ✅
• --ctx-size 4096 → 625 MiB free ✅ ← Production sweet spot
• --ctx-size 5120 → 623 MiB free ✅ (max safe)
• --ctx-size 6144 → 317 MiB free ⚠️ (too tight)
• --ctx-size 8192 → 179 MiB free ❌ (OOM risk)

Note: The Mamba-2 hybrid architecture means context barely affects VRAM up to ~5K. The cliff happens around 6K where compute buffers resize.

systemd Service (Auto-start on Boot)

[Unit]
Description=Nemotron IQ4_XS llama-server (CUDA)
After=network.target

[Service]
Type=simple
User=your-user
ExecStart=/opt/llama-server/llama-server \
  --model /path/to/nvidia_Nemotron-3-Nano-30B-A3B-IQ4_XS.gguf \
  --n-gpu-layers 28 \
  --cache-type-k q8_0 \
  --ctx-size 4096 \
  --parallel 1 \
  --flash-attn on \
  --port 8081 \
  --host 127.0.0.1
Restart=on-failure
RestartSec=5
Environment=CUDA_VISIBLE_DEVICES=0

[Install]
WantedBy=multi-user.target

OpenAI-Compatible API

Both Ollama and llama-server expose /v1/chat/completions. Drop-in compatible with any OpenAI SDK client:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8081/v1", api_key="none")
response = client.chat.completions.create(
    model="nemotron",
    messages=[{"role": "user", "content": "Hello!"}]
)

GPU Layer Guide

VRAM	GPU Layers	Est. tok/s	Notes
8 GB	20–22	~12	Tight, reduce context
11 GB	28	22–27	Sweet spot (this config)
16 GB	35–40	~28–32	Comfortable headroom
22 GB	52 (all)	~35–40	Full GPU offload 🚀
24 GB	52 (all)	~35–40	Full offload + large context

22GB upgrade note: With 22GB VRAM, all 52 layers fit on GPU — zero CPU offload, maximum throughput. Expected ~2x improvement over 28-layer config.

File Listing

File	Size	Description
nvidia_Nemotron-3-Nano-30B-A3B-IQ4_XS.gguf	16.8 GB	IQ4_XS quantized model (imatrix)
Modelfile	<1 KB	Production Ollama Modelfile
README.md	—	This file

Quantization Source

GGUF quantized by bartowski using llama.cpp release b7423 with imatrix calibration data. See bartowski/nvidia_Nemotron-3-Nano-30B-A3B-GGUF for all available quant sizes.

Credits

• Model: NVIDIA — Nemotron-3-Nano-30B-A3B
• Quantization: bartowski — IQ4_XS imatrix GGUF
• Benchmarking & Deployment: Tinker-Stack — Production validation on RTX 2080 Ti with Disclaw multi-agent war room