Step 5 Preview vs Nemotron 3.5 Super: Two Weight Drops, Two Very Different Licences

On 15 October 2026, StepFun published the checkpoints for Step 5 Preview and NVIDIA published the checkpoints for Nemotron 3.5 Super — the 120-billion-parameter, 12-billion-active vision-language model whose full name on the leaderboards is Nemotron 3.5 Super VL. Two flagship open-weight releases, one day, and almost nothing else in common. One is a 600B sparse mixture-of-experts built for long-horizon agent work and licensed under StepFun's own terms. The other is a hybrid Mamba-Transformer that runs at roughly two and a half times the output speed on a fifth of the active compute, under a licence designed to be used commercially without a conversation. On the index Artificial Analysis is serving today they read 29 and 14.

That score gap is the least interesting difference between them. The more consequential one is on the licence page, and it decides which of these two you can put inside a product without asking anyone's permission first.

15 October, twice

The two release histories are unusually parallel, which is why this comparison keeps being framed as a race. It is not one — the models were built for different jobs and only the shipping dates coincide.

• Step 5 Preview — API live 20 September 2026; BF16 and FP8 checkpoints published 15 October. StepFun's own announcement framed the delay as a deliberate one: ship the endpoint, gather traffic, then open the weights.

• Nemotron 3.5 Super VL — weights published 15 October 2026 as BF16 and NVFP4. NVIDIA shipped the checkpoints, the training recipes, the datasets and draft models for speculative decoding in the same drop, and its own technical report claims the model beats Nemotron 3 Ultra on SWE-bench Pro with roughly 4.6 times fewer active parameters.

• Distinct from its sibling — Nemotron 3.5 Lightning, the smaller text-only model released in the same drop, is a different product and not the subject here. Nor is this Nemotron 3 Ultra, which we compared against Step 5 Preview separately and which sits at a different point on the curve.

The licence is the bigger difference

Both are called open weights. Only one of the two licences is written to be used without a threshold.

• Step 5 Preview — the StepFun Model License 1.0. It reads like an Apache-style permissive grant with attribution, plus commercial conditions that trigger for very large deployments: the published terms attach separate-licence obligations once a product passes roughly 100 million monthly active users or $20 million in monthly revenue.

• Nemotron 3.5 Super VL — OpenMDW-1.1, which permits commercial use outright. No user-count tripwire, no revenue trigger, no separate negotiation.

• Weight formats — Step 5 Preview ships BF16 and FP8. Nemotron 3.5 Super VL ships BF16 and NVFP4, the four-bit format built for Blackwell hardware, with NVIDIA reporting 99.62% median accuracy recovery against the BF16 baseline across 36 benchmarks.

• What travels with the weights — Step 5 Preview published a technical blog describing the architecture and the training approach. NVIDIA published datasets, recipes and a drafter alongside the checkpoint.

For most teams the practical reading is simple: if you are under the StepFun thresholds — and almost everyone is — the two licences behave similarly in daily use. The difference shows up in diligence. A legal review of Nemotron 3.5 Super VL ends at "commercial use permitted". A legal review of Step 5 Preview ends with a paragraph about thresholds that someone has to model against your growth plan. That is a real cost, paid once, and worth knowing before the engineering work starts.

The board, after the re-score

The single most useful correction available here is that the widely quoted figures for both models are out of date. Launch coverage put Step 5 Preview at 44 and Nemotron 3.5 Super VL at 26. Artificial Analysis has since replaced the index version, and on the current board the same two models read 29 and 14. Everything derived from the older readings has moved with them.

Same run, one line per dimension:

• Intelligence Index — Step 5 Preview 29 vs Nemotron 3.5 Super VL 14.

• AA-Briefcase v1.1 — 1,423 vs 999.

• AutomationBench-AA — 51% vs 15%.

• Terminal-Bench 4.0 — 33% vs 0%.

• FrontierCode 2.0 — 12% vs 6%.

• AA-Omniscience — 16 vs −3, where the negative figure indicates more confident wrong answers than abstentions.

• Humanity's Last Exam (Diamond) — 24% vs 9%.

• Output speed — 179.1 vs 450.6 tokens per second.

• Cost per index task — $2.35 vs $0.56.

• Blended price per million tokens — $0.54 vs $0.1375.

The shape of the result is consistent across every sub-benchmark: Step 5 Preview leads on capability, Nemotron 3.5 Super VL leads on throughput and price. The Terminal-Bench 4.0 reading is the one that deserves a second look, because it is a zero, and zeros on a benchmark this widely used usually mean something specific rather than a total inability to perform the task.

NVIDIA's own numbers point somewhere else

NVIDIA's published figures describe a much stronger model than the independent run found. The company reports 51.78% on SWE-bench Pro, "up to five times" the agentic throughput of comparable open models, and 4,632 to 4,856 output tokens per second per GPU on GB200 hardware. Those are vendor-reported, produced on NVIDIA's own harness, and they are not the same tests that produced the index scores above.

This is not a contradiction, and treating it as one is the common mistake in coverage of this pair. SWE-bench Pro and Terminal-Bench 4.0 measure different things on different scaffolding, and a vendor harness tuned for its own model will tend to flatter it. The honest summary is that NVIDIA has published the most complete disclosure package of the two — recipes, data, a drafter — and simultaneously the least independently corroborated performance claim. Step 5 Preview is the reverse: a thinner disclosure package around a model whose behaviour independent labs have actually measured.

If you are choosing between these two on benchmark reputation, note which side of that trade you are comfortable with. A recipe you can re-run is worth more than a score you cannot.

What it costs to run each yourself

The parameter counts flatter Nemotron in a way that the memory bill does not fully repeat, but the direction is right.

• Step 5 Preview — 600B total, 27B active. The FP8 build fits a single node of four B300s or eight H200s. The BF16 build needs roughly 1.2 TB of GPU memory, which means eight B300s or two eight-GPU H200 nodes.

• Nemotron 3.5 Super VL — 120B total, 12B active, shipped additionally in NVFP4. It is a fifth of the total parameter count of Step 5 Preview, so the same hardware budget buys either a much smaller cluster or much more concurrency.

Active parameters decide inference cost; total parameters decide whether you can load the thing at all. Step 5 Preview's 27B active count keeps per-token compute competitive with models far larger than it, but the 600B envelope is what you have to fit in memory before any of that matters. Nemotron's 12B active count and 120B envelope are both small enough that a single node is a realistic deployment rather than a procurement project.

At NVIDIA's listed $0.225 in and $0.575 out per million tokens — against Step 5 Preview's $1.00 and $2.70 — the hosted route is cheap enough that self-hosting Nemotron 3.5 Super VL is mostly a data-residency decision, not a cost one. For Step 5 Preview, self-hosting is the only way to avoid a rate that is four to five times higher per token.

Routing the rest of your stack

Neither of these two models is in OrcaRouter's catalogue today, so this is not a piece where we can offer you a cheaper endpoint for either. What a router is genuinely useful for in this decision is the part that surrounds it: the 200-plus models that are in the catalogue are billed at provider list price with no markup, so a vendor price cut reaches your invoice the same day, and automatic failover means the non-Step-5, non-Nemotron traffic in your stack keeps running when one provider wobbles.

The realistic arrangement for a team that picks one of these is a hybrid: self-host the open weights for the volume workload, keep a hosted model behind the routing DSL for the tail cases where your cluster is saturated or the request needs a capability you did not deploy. OrcaRouter's fusion mode, which runs a panel of models and returns a combined answer, is a way to use that tail deliberately rather than as an emergency fallback.

What to watch next

Three things would change the answer here. If a second independent lab publishes Nemotron 3.5 Super VL results that resemble NVIDIA's own SWE-bench figures rather than the index run, the 29-versus-14 gap stops being the headline and the licence becomes the only real differentiator. If StepFun relaxes or clarifies the commercial thresholds in its licence, the diligence cost on the other side disappears. And if either model's hosted price moves, the four-to-five-times token gap is the number most exposed — Nemotron's list price is already low enough that a further cut would make self-hosting it look like a hobby rather than a strategy.

Until one of those happens, the honest split is this: Nemotron 3.5 Super VL is the model to deploy when you need throughput, a commercial licence without conditions, and a footprint that fits one node. Step 5 Preview is the model to deploy when the task is long, the context is large, and a wrong answer costs more than the tokens — and when you can live with a licence that asks a question at scale.