Run Nemotron Super 49B on a DGX Spark: reported tok/s
Nemotron Super 49B v1.5 at NVFP4 runs on one Spark under vLLM, but too slowly for one person: 5.79 tok/s single-stream in a published benchmark. It is a dense model, so every token reads all 49B parameters, and it only earns its place here as a batch model at high concurrency (695.1 tok/s aggregate at 256 requests). It is not a console preset; the command below installs it on a base image.
Measured on one Spark
| Engine | Quantization | Decode, single user | Prefill |
|---|---|---|---|
| vLLM | NVFP4 | 5.79 tok/s | not reported |
| Concurrency | Aggregate throughput | Per-request speed |
|---|---|---|
| 1 | 5.79 tok/s | 5.79 tok/s |
| 256 | 695.1 tok/s (advantage) | 2.85 tok/s |
Single-stream and concurrency figures from Dendro Logic's concurrency benchmark, running nvcr.io/nvidia/vllm:26.03-py3 on a single GB10 with 128 GB, on a character-extraction workload of roughly 1,500 prompt tokens and up to 400 output tokens. Both are quoted on GPUwerk's benchmarks page; neither source states an exact run date.
Memory footprint
Not measured. Neither GPUwerk's benchmarks page nor its models page states a weight size in gigabytes for Nemotron Super 49B v1.5 at NVFP4, so no footprint figure is given here. What both pages do establish is that it is a dense 49B model, meaning decode reads the full parameter count on every token, which is the direct cause of the 5.79 tok/s single-user number.
How to run it
Deploy a Spark with a base image (DGX OS or Ubuntu 24.04) and open the browser terminal; no SSH key is needed. Serve the NVFP4 build with vLLM's CUDA 13 container, following the vLLM doc. Because this model only makes sense batched, raise --max-num-seqs above the interactive default of 4 rather than leaving it low:
docker run -d --name vllm \
--gpus all --ipc=host --restart unless-stopped \
-p 8000:8000 \
-e HF_TOKEN="$HF_TOKEN" \
-v ~/.cache/huggingface:/root/.cache/huggingface \
vllm/vllm-openai:cu130-nightly \
vllm serve nvidia/Nemotron-Super-49B-v1.5-NVFP4 \
--gpu-memory-utilization 0.85 \
--max-model-len 32768 \
--max-num-seqs 64
Sweep vllm bench serve --num-prompts --request-rate against your own prompt shape before committing to a concurrency level, per the vLLM doc's verify-and-measure guidance.
What it is good for on a Spark
Not a chat model on this hardware. 5.79 tok/s is well under the roughly 10 tok/s line where a chat window feels usable, and no flag or quantization change fixes a dense model's bandwidth ceiling.
It is a batch model. At 256 concurrent requests, aggregate throughput reaches 695.1 tok/s, a 120-fold rise over the single-user figure, which suits overnight document processing, batch classification or an agent fleet where no single request needs to return quickly.
Read the per-request column before planning around the aggregate number: at that same concurrency, each request crawls at 2.85 tok/s. Find the concurrency level for your own workload that keeps per-request speed tolerable while capturing most of the aggregate gain, rather than assuming 256 is the right number for you.
Try it
A Spark is $0.79/hour, billed per minute against a prepaid balance (see the pricing page). Run your own concurrency sweep before deciding whether 5.79 tok/s single-user and 695.1 tok/s aggregate fit your batch workload. GPUwerk's first engagement covers picking and tuning a concurrency level for you instead.
Comparing with your own GPU or Mac? FitLocally's page for Llama 3.3 Nemotron Super 49B lists which graphics cards, Macs and mini PCs fit it at each quantization, with estimated speeds. FitLocally is a separate site I publish; its DGX Spark figures come from these guides.
FAQ
Is Nemotron Super 49B fast enough for a chat interface on a Spark?
No. It measures 5.79 tok/s single user, well under the roughly 10 tok/s line where a chat window starts to feel usable. It is a dense model, so every token reads all 49B parameters, and no configuration fixes that on this hardware.
What throughput does Nemotron Super 49B reach at high concurrency?
695.1 tok/s aggregate at 256 concurrent requests under vLLM, a 120-fold rise from the single-user number. Per-request speed at that concurrency is 2.85 tok/s, so this is a throughput result for batch work, not a latency one.
How much memory does Nemotron Super 49B need on a Spark?
Not measured. Neither GPUwerk's benchmarks page nor its models page states a weight size in gigabytes for this NVFP4 build, so no footprint number is given here rather than an estimate.