Nemotron 3.5 Super VL vs Nemotron 3.5 Lightning: Two Tiers of One Family, and the Token That Decides Which

These two models are not rivals, and treating them as one is the mistake the spec sheets invite. NVIDIA Nemotron 3.5 Super VL is a 120B-parameter, 12B-active vision-language model that needs four H100s or two GB200s to run at full precision. NVIDIA Nemotron 3.5 Lightning 30B-A3B is a 30B-parameter, 3B-active text model that fits on a single DGX Spark or one H100. They share a licence — OpenMDW-1.1, commercial use, open weights and open recipes — and they share a design philosophy of hushing per-token compute with Mamba-2 and sparse experts. They do not share an input modality, a hardware floor, or a job. Lightning is the older of the two by about two months, having shipped on 11 August 2026; Super VL's weights landed on 12 October 2026 with NVIDIA's announcement following on 15 October, so one of these is a settled release and the other is a day old.

The gate, in plain terms

Both models exist as downloadable checkpoints with published evaluation recipes. Both are vendor-evaluated: every number in the tables below comes from NVIDIA's own NeMo Gym or NeMo Evaluator harness, not from a third party, and neither has an independent throughput or accuracy audit behind it. That caveat applies to the whole comparison and is not repeated on every line.

Where Lightning sits better: NVIDIA says the Lightning 30B-A3B NVFP4 checkpoint has been pulled about 907,000 times on Hugging Face, against roughly 142 for the BF16 Super VL checkpoint at one day old. That is not a quality signal — it is the difference between a model you can run on a laptop-class accelerator and one you cannot.

What the two checkpoints contain

Nemotron 3.5 Lightning 30B-A3B is a hybrid mixture-of-experts built from interleaved Mamba-2 and MoE layers with a few attention layers, 30B total parameters and 3B active per token, 1M tokens of context, and text as its only input. NVIDIA trained it on more than 20 trillion tokens with an NVFP4 pre-training recipe, added Multi-Token Prediction heads, then fine-tuned on synthetic code, math, science, tool-calling and instruction-following data before a GRPO reinforcement-learning stage. It ships with three separate speculative-decoding strategies — DSpark for low-concurrency data-centre and DGX Spark work, plus MTP and DFlash — which is an unusual amount of inference engineering for a 30B model, and it tells you who the intended user is: someone running agents locally and watching the wall clock.

NVIDIA Nemotron 3.5 Super VL shares the skeleton and adds eyes. Same interleaved Mamba-2, MoE and attention hybrid, but with 512 routed experts per MoE layer, a C-RADIO vision encoder and a vision-language connector, 120B total and 12B active parameters, and text, image and video input. Its language backbone was initialised from Nemotron 3 Super Base, then distilled for a trillion tokens from Nemotron 3 Ultra Base before continued pre-training, a 35-billion-token extension to 1M context, software-engineering mid-training and a long-context recovery stage. The two releases also share a technical report — NVIDIA published the Super and Lightning work together.

• Active parameters — 12B per token for Super VL against 3B for Lightning. Four times the compute on every token generated, before the vision encoder is counted.

• Modalities — text, images and sampled video frames into Super VL; text in, text out for Lightning. Lightning's supported languages are English, Spanish, French, German, Italian and Japanese.

• Hardware floor — 4×H100 or 2×GB200 at BF16 for Super VL, with an NVFP4 build and a NIM microservice for production; one DGX Spark, one H100, an RTX 5090 or Ampere-class silicon via W4A16 for Lightning.

• Context — 1M tokens on both, both with published recipes rather than a single advertised maximum.

• Speculative decoding — Super VL uses its own MTP head natively, no drafter download needed; Lightning ships DSpark, MTP and DFlash variants as separate checkpoints.

• Licence — OpenMDW-1.1 for both, commercial use permitted, with open datasets and RL environments on the Hub.

What the numbers do and do not say

On NVIDIA's own harness the two are not close, and that is expected: a 12B-active model should beat a 3B-active one. Grounding the gap in NVIDIA's published tables: GPQA-Diamond 89.52 for Super VL against 75.44 for Lightning; HLE text-only 31.79 against 11.72; SWE-bench Verified 71.10 against 51.56; Terminal-Bench 2.1 66.70 against 24.58; AA-LCR 68.62 against 52.00.

The pattern inside those rows is more useful than the rows themselves. Super VL's lead is enormous on the long-context and agentic-execution benchmarks and comparatively modest on general knowledge — MMLU-Pro is 76.91 against Lightning's 81.94, which is a rare case of the smaller model winning a row. That is the shape of a model that was deliberately mid-trained on software-engineering data and given a long-context recovery stage: it is better at carrying a task through than at recalling facts. If your workload is factual question answering over short text, the 30B model's disadvantage mostly evaporates, and it does so at a quarter of the per-token compute.

NVIDIA lists a caveat on the Lightning card that is worth repeating verbatim in substance: the numbers were measured with and apply to the official NVFP4 checkpoint, and NVIDIA notes its own evaluation may differ from numbers the vendors of competing models report. That is an unusually honest line, and it applies equally to Super VL's table.

Picking the tier, and what happens above it

The decision rule falls out of the modality and the floor rather than the index. Anything that has to read a page — an invoice, a chart, a scanned form, a screenshot of an application, a video frame — can only go to Super VL, because Lightning has no path for pixels at all. Anything that is text in and text out, runs continuously, and lives on hardware you already own, belongs on Lightning, because a 3B-active model at NVFP4 on a single accelerator is a completely different cost base from twelve billion active parameters across four GPUs.

Above both of them sits NVIDIA Nemotron 3 Ultra 550B-A55B, which the Super VL card benchmarks against directly and which needs eight GB200s or sixteen H100s. Three tiers, three hardware classes, one licence. The family is designed so that a single organisation can deploy all three and move work between them, and the interesting engineering problem is not picking one — it is the router that decides which request goes where.

That routing layer is where a platform rather than a checkpoint earns its keep, and it is worth being precise about what we do and do not do here. OrcaRouter does not serve NVIDIA Nemotron 3.5 Super VL, Nemotron 3.5 Lightning 30B-A3B or Nemotron 3 Ultra 550B-A55B; none of the three is on our catalogue, and a self-hosted checkpoint is by definition not something an API provider can hand you. What we supply is the part that surrounds a mixed local-and-hosted fleet: the routing DSL that composes several models into a single call, so a pipeline can send its mechanical pages to a small cheap model and escalate only the hard ones, and automatic failover across upstream providers so that a hosted dependency degrading does not take the pipeline down with it. The catalogue side of that equation — currently more than two hundred models behind one key at 0% markup over provider list price — is what the escalation branch calls into. A self-hosted Lightning handling the volume and a routed hosted model handling the exceptions is a smaller and safer system than either half alone.

What to watch

Lightning's relative maturity is the underrated variable. It has been public since August, it has third-party quantisations, llama.cpp, Ollama, LM Studio and vLLM support, and a large enough download base that the failure modes are known. Super VL has a model card, a technical report, a vLLM day-zero recipe and a day-old download count. Running it in production this month means being the person who finds the rough edges.

The specific thing to watch is whether the NVFP4 Super VL build closes the hardware gap enough to matter. If a quantised 120B-A12B model runs usefully on a small number of accelerators — not one, but fewer than four — then the tier boundary moves and the whole three-model family becomes cheaper to deploy as a unit. Until that happens, the split is by modality: Lightning for text at volume on your own hardware, Super VL for anything with a pixel in it, Ultra when the task genuinely needs the largest model NVIDIA sells.