Nemotron 3.5 Super VL: which vLLM recipe to use on DGX Spark, 2× H100, or 1× GB200 — and why the card pins v0.32.0

NVIDIA announced Nemotron 3.5 Super VL on October 15, 2026 alongside the edge Bolt models. We already covered Bolt 74B-A9B in 24GB. This note answers the other half: how do I serve the 120B Super VL checkpoint, and which container?

Short version: use the NVFP4 weights for deployment, pin vllm/vllm-openai:v0.32.0 (or the card’s listed nightly digest), and copy the hardware-specific vllm serve block — Spark is not the same flags as 2× H100.

What Super VL is

From the NVFP4 model card and BF16 sibling:

Spec	Value
Parameters	120B total, 12B active
Architecture	Mamba-2–Transformer hybrid MoE + C-RADIO vision encoder
Context	Up to 1M tokens (deployment recipes may use less)
Inputs / output	Text, images, video / text
License	OpenMDW-1.1
NVFP4 deployment targets	1× DGX Spark, 1× H200, 2× H100, or 1× GB200
BF16 deployment targets	4× H100 or 2× GB200

vLLM 0.32.0 matters here because the card’s Quick Start explicitly names vllm/vllm-openai:v0.32.0 (plus a pinned nightly digest). That release’s headline fix is a Mamba2 chunk-scan NaN on SM90 (PR #61718) — relevant for this hybrid Mamba-2 stack.

Pick the recipe by box

Set:

export MODEL=nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4
export DRAFT_MODEL=nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr

DGX Spark (GB10) only — the card’s Quick Start uses LiLiCorr speculative decoding, --max-num-seqs 1, and seven draft tokens:

vllm serve $MODEL \
  --served-model-name nemotron-3.5-super-vl \
  --max-num-seqs 1 \
  --trust-remote-code \
  --gpu-memory-utilization 0.75 \
  --max-num-batched-tokens 8192 \
  --speculative_config.model $DRAFT_MODEL \
  --speculative_config.num_speculative_tokens 7 \
  --speculative_config.method dflash \
  --mamba-backend flashinfer \
  --mamba-ssm-cache-dtype float16 \
  --enable-mamba-cache-stochastic-rounding \
  --mamba-cache-philox-rounds 5 \
  --enable-mamba-shared-prefix-checkpoint \
  --video-pruning-rate 0.5 \
  --media-io-kwargs '{"video": {"fps": 2, "num_frames": 256}}' \
  --block-size 64 \
  --prefix-match-unit 64 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser nemotron_v3

2× H100 (up to 1M context) — native MTP, three speculative tokens, TP2 (one H200: same command with --tensor-parallel-size 1):

vllm serve $MODEL \
  --served-model-name nemotron-3.5-super-vl \
  --tensor-parallel-size 2 \
  --max-num-seqs 128 \
  --trust-remote-code \
  --gpu-memory-utilization 0.89 \
  --max-num-batched-tokens 16384 \
  --block-size 64 \
  --prefix-match-unit 64 \
  --compilation_config.max_cudagraph_capture_size 256 \
  --mamba-backend flashinfer \
  --mamba-ssm-cache-dtype float16 \
  --enable-mamba-cache-stochastic-rounding \
  --mamba-cache-philox-rounds 5 \
  --enable-mamba-shared-prefix-checkpoint \
  --mamba-cache-mode align \
  --speculative-config.method mtp \
  --speculative-config.num_speculative_tokens 3 \
  --video-pruning-rate 0.5 \
  --media-io-kwargs '{"video": {"fps": 2, "num_frames": 256}}' \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser nemotron_v3

Health-check with curl --fail http://localhost:8000/v1/models. API examples on the card use served name nemotron-3.5-super-vl and chat_template_kwargs.enable_thinking for reasoning on/off (plus medium_effort for the medium budget).

Local / edge vs this post

The NVFP4 card also points to a llama.cpp collection and Unsloth docs for DGX Spark / RTX Spark / RTX PRO 6000. For the compressed edge sibling that fits ~24GB, use our Bolt 74B-A9B llama.cpp recipes. BF16 Super VL is for customization and full-precision eval on multi-GPU boxes, not a laptop default.

Related

• Edge sibling: Nemotron 3.5 Bolt 74B-A9B 24GB llama.cpp recipes
• RTX Spark laptop memory tiers: which Spark memory tier to buy
• Another Mamba-related vLLM/llama.cpp path: NVIDIA Dory 4B looped LLM