Skip to content
trending

AI Tools

Nemotron 3.5 Super VL vs Step 5: Weights, Licenses and Test

Two October releases with different deployment needs and license terms. Nemotron completed all 12 Polish tasks; Step returned three empty answers.

By 6 min read
Nemotron 3.5 Super VL vs Step 5: Weights, Licenses and Test

👁 120 reads

NVIDIA Nemotron 3.5 Super VL and StepFun Step 5 Preview offer two different routes to running large models with downloadable weights as of October 2026. Both received the same 12 tasks in my test in Polish on October 19 via the API. Nemotron completed every task; Step completed nine and returned three empty answers at the response limit.

In short

  • Nemotron Super VL has 120B total parameters and 12B active parameters.
  • StepFun describes Step 5 as 600B total and 27B active.
  • NVIDIA uses OpenMDW 1.1; StepFun uses a custom license with commercial conditions.
  • The text-only Polish test scored 12/12 for Nemotron and 9/12 for Step.

What the two releases offer

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026. StepFun released Step 5 Preview weights on the same date, following the model’s earlier API availability. These are different release stories: NVIDIA expanded its model family, while StepFun made an existing hosted model available for deployment with downloadable weights. My notes on the Step 5 Preview API launch and its comparison with Kimi K3, GLM and Qwen remain the background; the other two releases of that week are in my open-weight AI models overview.

Super VL combines language and visual processing. NVIDIA positions it for document understanding, coding, reasoning, and tool use. StepFun emphasizes extended agent work, software engineering, and professional tasks. Both advertise long context, but those descriptions are starting points for evaluation rather than evidence about a particular user’s workflow.

Feature Nemotron 3.5 Super VL Step 5 Preview
Total and active parameters 120B / 12B 600B / 27B, headline figures
Advertised context Up to 1M tokens Up to 1M tokens
License OpenMDW 1.1 StepFun Model License 1.0
Polish test 12/12 9/12

StepFun’s card uses 600B while Hugging Face metadata reports 604B. The headline figure is useful for identifying the model, but it should not be mistaken for an exact inventory of every stored tensor. Similarly, active parameters describe computation per token, not the size of the complete weights that a deployment must manage.

Nemotron’s visual focus and deployment options

NVIDIA’s card describes a hybrid architecture combining Mamba-2, attention, and mixture-of-experts components with a vision encoder. The inputs include text, images, and video, with text output. Training resources and recipes accompany the broader release, giving developers material for customization as well as inference.

The NVFP4 card lists one DGX Spark, one H200, two H100 GPUs, or one GB200 as deployment options. The BF16 card lists four H100 GPUs or two GB200 systems. These examples refer to different precision formats and configurations. They should not be collapsed into a single “minimum hardware” number.

For a document assistant, the visual capability could matter when the input is a page image, chart, or screenshot. My invoice task supplied a written description, so Nemotron’s correct JSON answer was a text-extraction result. I did not test optical character recognition, video comprehension, or navigation of a visual interface.

It is also important to keep Super VL separate from the smaller Bolt models announced alongside it. Hardware claims for another model in the family do not transfer to Super VL. A familiar family name is not enough to establish whether a checkpoint fits a particular workstation.

Step 5 weights bring a larger infrastructure commitment

StepFun provides BF16 and FP8 checkpoints, and NVIDIA has also published an NVFP4 version. The official deployment examples use four B300 GPUs or eight H200 GPUs for FP8. BF16 examples use eight B300 GPUs or sixteen H200 GPUs. Those are substantial configurations even before considering the operational work of serving an application.

The documentation explicitly separates loading weights from allocating enough memory for a long request. Context storage and runtime buffers need additional capacity. Lowering context length or concurrency can be necessary even when the checkpoint itself fits. The one-million-token headline should therefore remain attached to a tested configuration, not assumed for every setup.

For a team interested in Step 5’s agent capabilities, the meaningful test is an entire workflow: interpreting the request, making valid tool calls, recovering from errors, and returning a useful result. Downloadable weights make that evaluation possible on controlled infrastructure. They do not demonstrate that it has already succeeded.

The licenses are not equivalent

OpenMDW 1.1 grants broad rights to the covered model materials. Redistribution requires preserving the agreement and applicable origin notices. It places no obligations on model outputs under that agreement, while retaining other provisions, including rights termination in specified litigation circumstances.

StepFun Model License 1.0 requires preserving notices and identifying modifications when distributing the model. Its current text, checked on October 20, 2026, also contains thresholds that matter:

  • If a commercial product using the model has more than 100 million active users or more than $20 million in revenue in any calendar month, it must display “Powered by StepFun AI” in its interface and documentation (section 2(b)).
  • Manufacturers or brand owners of hardware devices, such as phones, vehicles, wearables or robots, need a separate commercial license before shipping the model as a device feature, regardless of revenue (section 2(c)).
  • Companies running a model-as-a-service business need a separate commercial license once combined revenue exceeds $20 million over any consecutive 12 months (section 2(c)).

It should therefore not be described as having no revenue or user thresholds.

This difference is relevant before selecting a model for a commercial product. Internal research, distributing modified weights, selling model access, and embedding capabilities in hardware are distinct uses. The StepFun license defines conditions and exceptions for them. The license checked on October 20, 2026 is the basis for this description.

What the Polish test says about choosing between them

Each model received 12 prompts once, with temperature 0 and a 6,000-token completion allowance covering reasoning and the final answer. Nemotron returned accepted answers to all 12 (the full task list is in my test of four open models in Polish). Step returned accepted answers to nine, with empty responses on the autumn-formatting prompt, the administrative-region question, and the spelling correction.

Each Step failure recorded 6,000 completion tokens. Its longest request took 230.1 seconds, while the median was 2.15 seconds. Nemotron’s median was 1.20 seconds and its longest request took 11.7 seconds. These are elapsed times via the API, with potentially different serving backends, not measurements from equivalent local installations.

The saved usage records show about $0.0022 for Nemotron’s set and $0.0537 for Step’s set. Those are rounded historical charges. They do not establish current official list prices or predict the cost of a longer agent workflow. Step’s empty responses are particularly relevant because they consumed the full budget without producing usable text.

Nemotron’s score includes the disclosed spelling correction: “Weź parasol” was accepted as a valid alternative to the infinitive expected by the original grader. The test also tolerated some extra output formatting. Its scores should be read as accepted task answers rather than flawless compliance with every formatting instruction.

  • For short Polish tasks like these, Nemotron has the stronger observed completion result.
  • For visual workflows, evaluate Nemotron on actual images and documents before drawing conclusions.
  • For Step 5 agent workflows, test final-answer delivery under realistic reasoning budgets.
  • For either local deployment, evaluate the exact precision, runtime, and license that will be used.

FAQ

Is Step 5 bad at Polish?

This test does not establish that. It shows three empty responses under one budget and accepted answers on the remaining tasks. Larger samples and different settings could produce different results.

Can Super VL run on a consumer GPU?

The configurations cited here are NVIDIA’s published Super VL examples. They do not establish support for an ordinary 24 GB consumer card. Check the exact checkpoint and runtime rather than borrowing requirements from Bolt.

Which one should power an English application?

The Polish test cannot decide that. Use English prompts representative of the application and measure completion, correctness, format, latency, and cost together.

// read next

More on this from QUANTSLANT

Piotr Olszewski

Editorial team

Independent AI publication. Models and tools are tested on real tasks, prices are checked against official rate cards, and articles are updated when things change.

read nextHandling Unsolicited Recovery Offers After an Investment LossguidesHandling Unsolicited Recovery Offers After an Investment LossWhen an investment turns out to be fraudulent, the fallout rarely ends with the initial financial loss.

Most read

  1. 1Best AI Video Generator Apps for Your Phone (2026)
  2. 2Claude Opus 5.5 Motion Graphics: 60s Showreel from One Prompt
  3. 3Wan 2.2 on RTX 4090: timings, VRAM and four test clips
// site map
Quantslant in numbers 1,457posts in archive37news from the last 7 days18video models in the observatory
🎬 AI Video Benchmarks: Higgsfield DoP vs OpenAI Sora • Higgsfield vs Runway Gen-3 • AI Video Credit Calculator
× ‹ enlarge ›