Nemotron 3 Super: Updates Approaching Agentic Usability

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
, ,
1

Nemotron-3-Super-120B-A12B: Sub-second TTFT at 14.4 t/s on DGX Spark (llama.cpp, Q4_K_XL)

If you have been disheartened by 50 second TTFT, maybe we have been making the same mistakes,
or maybe todays (4/2/2026) update brought a welcome improvement. We are approaching usability on
a single DGX Spark for a model designed for agentic approaches. For me, this is a big deal. I have
heard predictions that we should be able to get 30+ tps on this when the inference stack gets
cleaned up a bit.

nVidia: there is this thing called Release Notes. You should try them sometime!

Hardware

  • NVIDIA DGX Spark (GB10)
  • ARM aarch64 — Cortex-X925 + Cortex-A725 (20 cores)
  • 128 GB unified memory
  • NVIDIA Driver 580.142 / CUDA 13.0
  • Ubuntu, kernel 6.17.0-1014-nvidia

Model

  • NVIDIA Nemotron-3-Super-120B-A12B (MoE — 120B total params, 12B active)
  • Quantization: Q4_K_XL (GGUF, ~78 GB on disk)
  • VRAM usage: ~84 GB

Inference Server

  • llama.cpp (build b8398-ee4801e5a)
  • Context: 16,384 tokens
  • 1 slot, no speculative decoding

Benchmark Results

Run Prompt TTFT Tokens TPS
1 Short (cold) 1,388 ms 113 14.2
2 Short (warm) 504 ms 116 14.6
3 Long output 814 ms 1,022 14.4
2

For Agentic scenario you should use vLLM or SGLang, the concurrency requirements are not the best for llama.cpp. Look how performance for a single Spark change as concurrency increases.