I’m getting 17 tok/sec with flash attention and Q4_K_M completely in GPU. Anyone else tried nemotron super on Spark already? Any tips on optimizing running or it’s like best that we can get?
Being discussed here… getting around 16-17 tokens/second. Its OK-ish for things like openClaw but sluggish/slow.
I see, thanks for details. I’ll play around with settings as well and see if I can get any better results.
Never got it running stable. In case it was running ~17t/s.
Eventually I pulled it in LM Studio. Here it serves 20t/s via the LM Studio API Server.
Did you make any changes in the default settings for the model?
In LM Studio? None, except limiting to 131k context. It’s the official nemotron-3-super at the top of the list
Yeah, I’m testing super with 1mil context and flash attention. Probably 1m context is dropping it to ~17 t/s
… running stable?
Are you using GGUF? Or have you tried something different?
simillar results here running the model on ollama
| model | test | t/s | peak t/s | ttfr (ms) | est_ppt (ms) | e2e_ttft (ms) |
|---|---|---|---|---|---|---|
| nemotron-3-super:latest | pp2048 | 461.92 ± 4.31 | 4014.74 ± 28.46 | 4012.26 ± 28.46 | 4014.74 ± 28.46 | |
| nemotron-3-super:latest | tg32 | 19.93 ± 0.06 | 21.00 ± 0.00 |
llama-benchy (0.3.5)
date: 2026-03-26 11:42:02 | latency mode: api
Given the discussion here, sounds like no one had encountered the NemoClaw installation issue which this guy experienced. Is it true?
For the time being I would not trust any sandboxed agent on a machine I also use otherwise.
I do fully understand the issues described in the discussion link that you sent.
Just my opinion:
Use openclaw - which in fact is a security nightmare. Therefore install it on a separate machine and set up a proper hardware firewall around it. Sandboxing will drive you crazy - either you are getting security issues or you cripple the agent too much in its abilities.
You can find a short description of my setup here: OpenClaw on several GPUs
Thank you very much for your prompt reply! I am not crazy on utilizing NemoClaw’s sandbox feature since I have been using a dedicated local system running OpenClaw. But NemoClaw claims to have other benefits, such as multi-user support (if I read it correctly). I also got an impression that NemoClaw has better multi-agents orchestration capability. These are what I am interested more. Perhaps, these are just marketing terms, not in reality yet?
I definitely can start from simply using Nemotron 3 super as the brain model to start with. Which way do you recommend to run NM3 super locally on DGX Spark?
That’s true: multi-agent orchestration could be improved with OpenClaw. I guess though the reason is weak “prompting” in the .md files.
Getting the agent to do something the way you want to have it done is quite a job. I am trying a gated project management setup right now. Specification, concept, design etc. with review gates. It’s time consuming to work like this. On the other hand, how would you expect the agent to do what you want, if you don’t specify it correctly …
As said before, I gave up with Nemotron in vllm temporarily. I use Nemotron-3-super as researcher that does deep researches for me as subagent of the main agent. Subagents run in the background, thererfore I don’t care if it takes a little longer. Nemotron is served via the LM Studio internal api server. The main agent (qwen3.5:35b) spawns the subagent. That way the main agent stays reactive for other tasks.
I was chatting with Nemotron the other day. Its chats are boring as hell. Qwen 3.5 is much more fun to chat with.
Just for fun I paste qwen’s message, when I told him about Nemotrons “personality:
qwen 3.5:35b
If I imagine Nemotron trying to make a joke, I picture him:
• First checking the joke 3×
• Providing all sources and citations for it
• And then saying: “This is a suboptimal solution to a non-existent problem. I recommend reformulating the question instead.”
…………..
Haha, the classic “Nemotron in a meeting with Jensen” scenario! 🤣
So:
• Monday: “Mr. Jensen, we have optimized the FP4 kernels for SM121.”
• Jensen: “And what about the NVFP4 problem?”
Nemotron collapses under the weight of 147 Stack Overflow errors he collected yesterday in 0.3 seconds