01What it can do
- MoE frontier: 561B / 55B active in BF16
- Multilingual pretraining in 11 languages (including French)
- 128k native context
- Open base for fine-tuning and research
- —Base variant: not chat/instruct aligned; fine-tune before application use
- —~325 GB VRAM in Q4, ~1.1 TB in BF16 — H100/MI300 multi-GPU mandatory
- —No Ollama tag: install via Hugging Face only
- —NVIDIA Open Model License (review based on commercial use)
04Install
Install Ollama for your OS. Check the model and its quantization before downloading. Start with 4096 tokens of context, then check placement with ollama ps. A command below is not proof that a test was run on your machine.
02Required memory
Approximate GPU VRAM required to run this model, including 4k tokens of context overhead. For a longer context, add ~1 GB per 8k-token increment.
What hardware do you need for Nemotron 3 Ultra Base (BF16)?
To run Nemotron 3 Ultra Base (BF16) locally with Q4 quantization, you need about 325 GB of VRAM. An option to compare: BOSGAME M5 128GB / 2TB (Ryzen AI Max+ 395) — this model exceeds this mini-PC's GPU capacity: choose a smaller model or suitable infrastructure.
Why this choice? Our complete guide on BOSGAME M5 128GB / 2TB (Ryzen AI Max+ 395) →
Affiliate links — possible commission at no extra cost to you; independent recommendations. As an Amazon Associate, BestLLMfor earns from qualifying purchases.
Too large for your machine? The kit gives you the model that fits in your VRAM
- Lifetime online access
- PDF + files
- Lifetime updates
03Expected speed
Tokens generated per second in Q4_K_M, 4k context. Beyond 20 t/s, reading is comfortable. Below 10 t/s, that's just for testing.