Local AI · Recipe · 1× DGX Spark

Nemotron-3.5-Super-VL 120B-A12B NVFP4 with vLLM on 1× DGX Spark

Nemotron-3.5-Super-VL 120B-A12B NVFP4 with vLLM on 1× DGX Spark: 48 tok/s according to github.com (dataset as of Nov 3, 2026).

by sfxnz

48tok/sMixed

1 request, mixed prompt set (SPEED-Bench), greedy, thinking off, with LiLiCorr (per recipe)

Source: github.com
Intelligence (original model) · Artificial Analysisno independent value

Install: Nemotron-3.5-Super-VL 120B-A12B NVFP4 with vLLM on 1× DGX Spark

Command
git clone https://github.com/sfxnz/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-DGX-SPARK.git && cd NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-DGX-SPARK && ./run.sh

Full instructions in the README
Repo

Engine

Engine
vLLM 0.32.0rc1 (nightly 3f1bcfd8, pinned vllm/vllm-openai@sha256:cb9bc64f…), LiLiCorr speculative decoding (dflash, 7 tokens), fastsafetensors
Quantization
NVFP4 experts, FP8 Mamba projections and attention, FP8 KV
Model family
Nemotron-3.5-Super-VL-120B-A12B
Context
1,048,576
Parameters
120B / 12B aktiv (Hybrid Mamba-2 + LatentMoE, 512 Experten, 22 aktiv)
Creator
sfxnz
GitHub stars
1
Repo updated
Oct 16, 2026

Measurements

Every sourced value of this recipe, each with its condition and source. Bars relative to the largest value in the group.

Decode · 1 request

  • Mixed (decode)48 tok/sVllm bench serve, 32 Prompts aus 369 SPEED-Bench-Prompts, greedy, Thinking aus, bis 512 Ausgabe-Tokens (Ø ca. 265), pro Nutzer, Mittel aus 2 Boots (50,4 / 45,3); TTFT p50 0,40 s. Karten-Sampling (T=1,0, top_p 0,95) mit Thinking: 38 Source 

Decode · parallel requests

Sums over all streams are marked; per request it is correspondingly less.

  • 32 parallel requests189 tok/sPeak (decode) · Ausgabe-tok/s aller Nutzer bei 32 Nutzern, Default-Profil (ca. 24 passen gleichzeitig in den Pool, TTFT p50 7,4 s); 24 Nutzer: 179 Source 

Prefill by context

  • 32K2,466 tok/sEin Request, Zufalls-Tokens; 8K 2.506 / 128K 2.042 tok/s Source 

Time to first token

Shorter is better.

  • Prompt not stated400 ms1 Nutzer, p50, SPEED-Bench-Prompts, greedy, Thinking aus Source 

What you need

Hardware
1 × NVIDIA DGX Spark (GB10)
Engine
vLLM 0.32.0rc1 (nightly 3f1bcfd8, pinned vllm/vllm-openai@sha256:cb9bc64f…), LiLiCorr speculative decoding (dflash, 7 tokens), fastsafetensors
Context
1,048,576 tokens

Related recipes

1× DGX Spark
71.8tok/sEveryday

1 request, prose prompt (sampled, thinking off), with LiLiCorr (per recipe) Source 

Intelligenceno independent value
  • ≤ 3 bit: quantization may cost quality
  • Custom kernel required
  • License unclear

24GB pack: IQ2_XXS experts (2.0625 bits), NVFP4 Mamba projections, Q8_0 attention/latent projections/lm_head; bf16 KVTensorFold tensorfold-native v1.1.6 (eea9dc9c) + recipe patches/ (Bolt lane backend, libbolt.so CUDA forward pass), LiLiCorr draftsWeights Repo Updated Nov 1, 2026

Details
1× DGX Spark
59.4tok/sEveryday

1 request, chat prompt (20 prompts, thinking off), no drafting (default profile) Source 

Intelligenceno independent value
  • Custom kernel required

EXL3 6.08 bpw (28,3 GB), FP16 KV; nur Textpfad (Vision-Tower wird von ExLlamaV3 nicht ausgeführt)ExLlamaV3 1.5.0 (Fork r0b0tlab/exllamav3-r0b0t, Branch minicpm-v4.7-dspark @ 5ddae7d12448), torch 2.14.0+cu130Weights Repo Updated Oct 14, 2026

Details
1× DGX Spark
55.6tok/sEveryday

1 request, prose prompt, with MTP (per recipe) Source 

Intelligence40with thinking
  • Custom kernel required
  • License unclear

NVIDIA FP4-QAD-Export (FP4-Local @b9ce18a6): Routed Experts NVFP4 (Gruppe 16), Attention/Shared Expert MXFP8, alle 512 Experten mit 10 pro Token; n-gram-Tabelle per pread von der SSD; FP8 KVTensorFold tensorfold-native (Zig) v1.1.4 (eab85828) + Patches 0001–0017, Image tensorfold-qwen38:v1.1.4-eab8582 (GHCR v1.1.4-eab8582-4baf16137ecd) auf nvcr.io/nvidia/pytorch:26.07-py3, MTP-DraftsWeights Repo Updated Oct 30, 2026

Details

← Back to overview

Local AI in your company?

In a workshop we work out which models and which hardware fit your tasks, and build the first agent on your infrastructure.