Development ·
NVIDIA releases Nemotron 3.5 weights and deployment guidance
The release adds Super vision-language and Lightning text models, with published weights, licensing and runtime guidance.
Why it matters
Teams can inspect and host these models, opening a concrete path to compare multimodal services with infrastructure they control.
What to watch
Independent results on the documented hardware, deployment costs under sustained load, and practical compatibility of the licensed artifacts.
Still uncertain. Hardware requirements vary by precision and modality. The published performance and resource claims were not independently reproduced.
Impact assessment reviewed 2026-10-19.
- Who it helps
- Teams evaluating self-hosted models and multimodal inference.
- Availability
- The inspected Super NVFP4 card lists downloadable weights under OpenMDW-1.1 and Linux runtime recipes. It lists DGX Spark, H100, H200 and GB200 configurations with different modality support.
Plan for the complete model
Super has 120 billion total parameters and activates about 12 billion per token; Lightning has 30 billion total and about 3 billion active. Active parameter counts describe computation and do not by themselves specify the memory needed to load a model.
Match the recipe to the job
The Super card distinguishes text/image and video-capable configurations. Check the exact quantization, runtime version and hardware row for your intended modality before allocating a machine.
What you can do
- Use the model card’s matching hardware recipe as the starting point for a small, fixed evaluation set, and record memory, latency and task success together. Open reference ↗
Where the evidence stops
No inference was run. Vendor benchmark comparisons and hardware recipes have not been independently reproduced here.
Sources read. Primary text checked October 19, 2026 UTC. The external product or research workflow was not executed; reported results remain attributed to their authors.
Sources & further reading
Web · official
Nemotron 3.5 Super and Lightning ↗
October 14 release page read for architecture, modalities and available artifacts. Performance comparisons are vendor reports.
NVIDIA · page checked 2026-10-19 · Published 2026-10-14
Hugging Face · implementation
Nemotron 3.5 Super VL NVFP4 model card ↗
Read license, supported runtime and hardware recipe matrix. No model downloaded or inference executed.
NVIDIA · page checked 2026-10-19
Coverage note: The linked sources establish the documented scope. Unlisted social platforms and independent execution remain unverified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
2026-10-14: The release adds Super vision-language and Lightning text models, with published weights, licensing and runtime guidance.
Assessment: Teams can inspect and host these models, opening a concrete path to compare multimodal services with infrastructure they control.