Nemotron 3 Ultra Base (BF16)

Overview

A frontier-scale MoE base checkpoint from NVIDIA: 561B total parameters with ~55B active, pretrained across 11 languages. Unaligned — built for fine-tuning and research, not direct chat use, and requires datacenter-class multi-GPU hardware.

When to pick this model

• You're a research team fine-tuning a frontier-scale multilingual MoE from scratch
• You have multi-GPU H100/MI300-class infrastructure and need an open base to align yourself
• Multilingual pretraining coverage (11 languages) is a requirement for your downstream model
• You need full control over alignment/instruct-tuning rather than a pre-aligned model

VRAM requirements by quantization

VRAM figures include model weights plus a typical 8k KV cache and ~600 MB runtime overhead (Ollama / llama.cpp baseline). Add headroom for higher context lengths.

In practice, Nemotron 3 Ultra Base (BF16) is server-class even at Q4_K_M (325 GB). Stepping up to Q8_0 nearly doubles the footprint to 600 GB, and unquantized FP16 weights take 1122 GB — plan your GPU around the Q4 or Q5 figure unless you specifically need the higher fidelity.

Without a GPU, Nemotron 3 Ultra Base (BF16) needs roughly 729 GB of system RAM to run on CPU via llama.cpp or Ollama — workable for background jobs, but far slower than GPU inference. Throughput estimates from our compatibility engine: around 1.5 tokens/sec on entry-level GPUs, on the order of 2.5 tokens/sec on a mid-range card, and up to 5 tokens/sec on high-end hardware — assuming the chosen quantization fully fits in VRAM.

What hardware do you need

The table below matches Nemotron 3 Ultra Base (BF16) to common GPU memory tiers, using the highest-fidelity quantization that fully fits each card class. Spilling layers to system RAM works but costs most of the speed, so size your card to the quantization you actually want to run.

Which GPU should you buy to run Nemotron 3 Ultra Base (BF16)?

To run Nemotron 3 Ultra Base (BF16) locally at Q4, you need ~325 GB of VRAM. The best value for this is a Apple Mac Studio (64+ GB unified memory).

Strengths

• Frontier-scale MoE: 561B total / ~55B active parameters in BF16
• Multilingual pretraining across 11 languages, including French
• 128K native context
• Open base weights suitable for custom fine-tuning and research

Limitations

• Base checkpoint only — unaligned, not usable for chat or instruction-following until fine-tuned
• ~325GB VRAM at Q4, ~1.1TB in BF16 — requires multi-GPU H100/MI300-class datacenter hardware
• No Ollama tag; HuggingFace-only distribution
• NVIDIA Open Model License terms should be reviewed for commercial use

Typical workloads

In our catalog grid, Nemotron 3 Ultra Base (BF16) is filed under Multilingual Fine-Tuning, MoE Research, Custom Alignment — the use cases where its size/quality trade-off makes the most sense. Its tags translate to concrete workloads: multilingual workloads.

The 125k-token context window covers long chats and mid-sized documents, though very large retrieval workloads will need chunking. It ships under the NVIDIA Open Model License license — commercial use is generally possible but read the specific terms before embedding it in a product.

Architecture & training

Architecture: Mixture-of-Experts · 561B total parameters · ~55B active per token · 128K context · BF16 weights · latent-MoE + MTP

Training: Base checkpoint (pretrained, non-instruct) from the Nemotron 3 Ultra family. Multilingual pretraining covering 11 languages (en, fr, es, it, de, pt, ja, ko, hi, ar, zh). Released in BF16 precision on Hugging Face.

Quick start

# HuggingFace : nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16

Or use the open-source MCP server to query this model from Claude Desktop, Cursor, or any MCP-compatible client.

Similar models worth comparing

Frequently asked questions

How much VRAM does Nemotron 3 Ultra Base (BF16) need?

At the recommended Q4_K_M quantization, Nemotron 3 Ultra Base (BF16) needs about 325 GB of VRAM. Q8_0 takes 600 GB, and unquantized FP16 weights take 1122 GB.

Can Nemotron 3 Ultra Base (BF16) run without a GPU?

Yes — with roughly 729 GB of system RAM it runs CPU-only through llama.cpp or Ollama. Expect a fraction of GPU speed, which is fine for background or batch jobs but slow for interactive chat.

What context window does Nemotron 3 Ultra Base (BF16) support?

Nemotron 3 Ultra Base (BF16) supports a 125k-token context window (128,000 tokens).

Can I use Nemotron 3 Ultra Base (BF16) commercially?

Nemotron 3 Ultra Base (BF16) ships under the NVIDIA Open Model License license. Commercial use is generally permitted subject to its terms — review the license text before shipping a product.

How fast is Nemotron 3 Ultra Base (BF16) on consumer hardware?

Our compatibility engine estimates on the order of 2.5 tokens/sec on a mid-range GPU and up to 5 tokens/sec on high-end cards, assuming the quantization fully fits in VRAM.

Which quantization of Nemotron 3 Ultra Base (BF16) should I download first?

Start with Q4_K_M (325 GB) — the standard size/quality sweet spot. Step up to Q5_K_M or Q8_0 only if you have VRAM headroom. It does not fit a single 24 GB consumer card — plan for multi-GPU or server hardware.