NVIDIA Nemotron 3.5
Nemotron 3.5 includes distinct NVIDIA checkpoints. Super VL is a 120B-total/12B-active vision-language model with up to 1M context. Bolt is a 74B-total/9B-active model; its 24GB variant uses native 2-bit compression and supports up to 256K context. These are separate checkpoints and deployment choices.
Inspect NVIDIA’s model collection ↗Why it matters
Choose by the exact checkpoint, available memory and task. A family name or small active parameter count alone does not identify a runnable configuration.
Super VL: long-context visual and document work
The Super VL card accepts text, images and video and returns text. It documents deployments with four H100 GPUs or two GB200 GPUs. The full context budget includes the prompt, visual content, tool results and generated response; usable capacity also depends on KV cache and runtime settings.
- Super VL contract
Exact checkpoint and deployment limits.
Bolt: distinguish BF16 from the packed 24GB variant
Bolt’s BF16 checkpoint and NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB are different artifacts. The packed card targets a footprint below 24GB at batch size one and 128K context. Its advertised 256K ceiling does not guarantee that every 24GB machine can hold that request plus runtime buffers.
- Bolt contract
Compare formats and memory conditions.
Licensing and supported runtimes
The cards identify OpenMDW-1.1 as the governing license and describe commercial and non-commercial use. Follow the selected card’s pinned runtime and parser instructions. Keep licensing, checkpoint availability and a successful local deployment as separate checks.
Release dates and Lightning’s separate history
NVIDIA’s project index dates the Super/Lightning collection to October 14, while the Super and Bolt Hugging Face cards identify October 15 releases. Lightning’s card records an earlier August 11 date. Do not read the collection announcement as proof that every member first shipped on the same day.
- NVIDIA project index
Read the collection dates and linked model cards.
What to test before adopting it
For OCR or document work, compare the answer against the actual page, table or image and record the crop, input resolution and prompt. For tool workflows, verify every action’s postcondition. NVIDIA’s examples and benchmarks are useful starting points; this guide does not report an independent visual-quality or hardware test.
Useful first jobs
- Super VL: long-context visual and document work
- Bolt: distinguish BF16 from the packed 24GB variant
- Licensing and supported runtimes
Frequently asked questions
Does 24GB mean every Bolt request fits in 24GB?
No. The packed-card target is batch size one at 128K context. Cache, runtime buffers and a larger request can require more memory.
Is Super VL the same checkpoint as Bolt?
No. Super VL has 120B total parameters; Bolt has 74B and separate numerical formats.
Sources and evidence
Checked October 21, 2026 against the linked primary sources. Specifications and examples are publisher documentation; no independent model or hardware run is claimed.
- Family project index ↗
NVIDIA · Primary source
- Super VL model card ↗
NVIDIA · Primary source
- Bolt packed model card ↗
NVIDIA · Primary source
- Bolt BF16 model card ↗
NVIDIA · Primary source