Yes, Nemotron 3.5 Super VL DGX Spark is a real pairing: NVIDIA optimised the model to run on its DGX Spark, and the 128GB version has enough memory to hold it locally. Per NVIDIA, LiLiCorr speculative decoding makes generation up to 2.3x faster on DGX Spark. But buying a DGX Spark just for this one model only makes sense if you need local, private AI every day.
📺 Watch: NVIDIA Nemotron 3.5 Super VL Is INSANE! (FREE AI Model)
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I'll explain what DGX Spark is, why this model fits it, what EXO Labs found when they tested it, and who should buy the hardware versus just using the model hosted for free.
If you want help setting up open models on your own machine or in the cloud, my AI Profit Boardroom ($69/mo) includes the Agent OS zip, a video tutorial and a 30-day roadmap.
What Is DGX Spark?
DGX Spark is NVIDIA's small desktop AI computer. It sits on your desk instead of in a data centre. The version EXO Labs used in the evaluation NVIDIA mentions has 128GB of memory, and that number is the whole reason it matters here.
Nemotron 3.5 Super VL has 120B total parameters. It's mixture-of-experts with 12B active per token, but all the weights still need to fit in memory. Local runs need roughly 70-90GB. A normal laptop can't do that. A 128GB DGX Spark can, with room left for images and context.
For the full background on the model itself, read my Nemotron 3.5 Super VL overview.
Why Nemotron 3.5 Super VL DGX Spark Works Well Together
NVIDIA built three things into this model that make it a good fit for desktop hardware.
1. NVFP4 quantization
The model ships in NVFP4, NVIDIA's 4-bit format, built using quantization-aware distillation. On Blackwell hardware it runs at 4-bit, which shrinks the memory footprint while keeping quality. On non-Blackwell hardware, per NVIDIA, the same checkpoint runs with 4-bit weights and 16-bit activations.
2. LiLiCorr speculative decoding
On DGX Spark, NVIDIA reports LiLiCorr speculative decoding gives up to 2.3x faster generation. Speculative decoding means the system guesses ahead and checks those guesses quickly, so you wait less for each answer.
3. Multi-token prediction
The model has an internal drafter that proposes several tokens at once, which get checked in one pass. That's part of why NVIDIA reports up to 5x higher single-GPU throughput on agentic tasks versus open models its size.
Put those together and you get a big model that actually feels usable on a desktop box, not just technically loadable.
📺 Watch: OpenClaw + Nvidia Nemotron 3 Super + Ollama is INSANE!
What EXO Labs Found
NVIDIA points to an independent evaluation by EXO Labs on a 128GB DGX Spark. EXO Labs put Nemotron 3.5 Super VL on the Pareto frontier for accuracy versus task time.
In plain English: among the models they tested on that machine, nothing gave you better accuracy without also taking longer, and nothing was faster without also being less accurate. It sits on the best-trade-off line.
That's useful, but keep it in perspective. It's one evaluation on one machine, and it's about trade-offs, not raw "best model" status. Most of the other numbers you'll see, like the 2.3x speed-up, are NVIDIA's own with no independent replication yet. For how the benchmark claims stack up, read my Nemotron 3.5 Super VL benchmark breakdown. And for how I test models on real jobs, see Goldie Bench.
📺 Watch: New FREE Nvidia Nemotron 3 Super + Ollama
How to Run It on DGX Spark
You've got the same two easy options as any big-memory machine.
Option 1: Ollama
- Install Ollama
- Run ollama run nemotron-3.5-super
- Wait for the roughly 80-90GB download
- Plug it into an agent with a launch command, for example ollama launch opencode --model nemotron-3.5-super
Ollama lists the model as nemotron-3.5-super, with vision, tools and thinking tags and a 1M context window. Hermes, Claude Code and OpenClaw all have launch commands too.
Option 2: LM Studio
- Install LM Studio
- Search for nvidia/nemotron-3.5-super-vl
- Download the Q4_K_M build, about 70GB
- Load it and attach an image to test the vision side
Either way, remember the 1M context window is a ceiling. Usable context depends on memory and runtime. Even on 128GB, very long prompts plus lots of images will eat your headroom fast. Start modest.
For builders who want to serve it properly, NVIDIA also lists vLLM, SGLang, NIM and Exo as options. Follow each tool's own docs for DGX Spark settings.
Should You Buy a DGX Spark for This?
This is the real question. Here's my honest take.
| Your situation | My recommendation |
|---|---|
| You want to try the model | Use it hosted. OpenRouter has nvidia/nemotron-3.5-super-vl with a :free variant. |
| You build agents and want a free model | OpenCode free. Pick "Nemotron 3.5 Super VL" in the drop-down. |
| You handle private client documents daily | Local hardware starts to make sense. |
| You want to fine-tune or run models all day with no rate limits | A 128GB DGX Spark or big-memory workstation is worth looking at. |
| You have a normal laptop and a normal budget | Don't buy hardware. Go hosted. |
The trap is buying expensive hardware for one free model. The model is free hosted. The hardware isn't. Only buy if you'd use local AI across many models and many jobs, not just this one.
If you're deciding whether hardware makes sense for your business, book a free AI strategy session and we'll look at your actual workload before you spend anything.
Good Jobs for a Local Desktop Setup
If you do go the DGX Spark route, these are the jobs where running locally actually pays off:
- Private documents: contracts, client reports and financial PDFs that you'd rather not send to a cloud API
- Screenshot to code: hand it a design screenshot and get working HTML back, as many times as you like
- Chart and dashboard reading: NVIDIA reports 80.8% on CharXiv chart reasoning, and 84% with Python code interpretation
- Always-on agents: Hermes or OpenClaw running all day without hitting free-tier limits
- Receipts and invoices: batch-reading them into lists without paying per call
If none of those sound like your daily work, that's your answer. Stay hosted.
What I Use It For
I've used Nemotron 3.5 Super VL free inside OpenCode, in my Agentic Operating System. I asked it to build a beautiful to-do list app, make it entertaining, engaging, and the build came out working. It also shows up as a model I can pick in Hermes Agent.
Honest take: it's not a frontier model and it's not going to replace Opus 5.5. But it's free, open, fast, and it reads screenshots, PDFs, charts and receipts. On a desktop box like DGX Spark, that means you can process private documents without them leaving your desk.
NVIDIA also released smaller Nemotron 3.5 Bolt edge models alongside it, including a 74B-A9B for 24GB systems and an E4B for 4GB edge hardware. If 128GB is out of reach, those are different, smaller models worth knowing about.
To compare how NVIDIA positions the whole release, read my NVIDIA Nemotron 3.5 Super VL breakdown. And to run several agents side by side on whatever hardware you have, my Agent OS guide shows how I set up Hermes, OpenClaw, OpenCode and Claude.
Inside the AI Profit Boardroom, 3,400+ members get 4 weekly coaching calls and daily tutorials. Or grab a free AI strategy session to talk it through one on one.
FAQ
Can Nemotron 3.5 Super VL run on a DGX Spark?
Yes. NVIDIA optimised it for DGX Spark, and EXO Labs evaluated it on a 128GB DGX Spark, where it landed on the accuracy versus task-time Pareto frontier.
How much faster is it on DGX Spark?
Per NVIDIA, LiLiCorr speculative decoding gives up to 2.3x faster generation on DGX Spark. That's NVIDIA's own figure.
Is DGX Spark worth buying just for this model?
For most people, no. The model is free hosted on OpenRouter and OpenCode. Buy hardware only if you need private, local AI every day across many jobs.
Does it need Blackwell?
NVFP4 runs at 4-bit on Blackwell. Per NVIDIA, the same checkpoint runs on non-Blackwell hardware with 4-bit weights and 16-bit activations.
