Summarize with AI
NVIDIA has released Nemotron 3.5 Super VL, an open-weights model that understands text, images and video up to two minutes long. It is a Mixture of Experts (MoE) model with 120B total and 12B active parameters, built for agents that run all the time. NVIDIA also released Nemotron 3.5 Bolt, a compressed 74B version that fits in 24 GB of memory. This article covers what the two models are, how they are built, how they score, and where you can run them.
Both models came out on October 15, 2026, under the OpenMDW-1.1 license, which allows commercial use.
What Super VL is for
NVIDIA describes Super VL as a model for "always-on agents": agents that follow instructions, use tools and work through multi-step business workflows.
On the vision side, the focus is documents:
- optical character recognition (reading text in images);
- tables and charts;
- web pages;
- PDFs and other documents.
Reasoning can be switched off or on through the chat template, and there is a medium-effort mode in between (enable_thinking plus medium_effort).
How it is built
| Part | Super VL |
|---|---|
| Parameters | 120B total, 12B active |
| Language backbone | Hybrid of Mamba-2 and Transformer layers with Mixture of Experts |
| Experts | 512 routed experts in each MoE layer |
| Vision encoder | C-RADIO |
| Faster generation | Multi-Token Prediction for native speculative decoding |
| Context | Up to 1M tokens |
| Video | mp4, up to 2 minutes |
Bolt: the 24 GB version
Nemotron 3.5 Bolt 74B-A9B is Super VL compressed to run on a single consumer-class memory budget.
- Fewer, narrower experts. NVIDIA's compression picks a different number of experts and a different expert width for each layer. The result has 74B total and 9B active parameters. Each token uses 9 experts, down from 22 in Super VL.
- Retrained to recover quality. After compression, the model was trained again with knowledge distillation on 120B tokens.
- 2-bit experts. All expert weights (routed and shared) are quantized to about 2 bits. Mamba layers stay in NVFP4, and embedding and output layers in 8-bit.
- 24 GB. The packed model fits in 24 GB of memory, with a context of up to 256K tokens.
Benchmarks
NVIDIA's model card compares Super VL with the closest model in size, Qwen3.5-122B-A10B. The two are nearly level:
| Benchmark | Nemotron 3.5 Super VL | Qwen3.5-122B-A10B |
|---|---|---|
| GPQA-Diamond | 89.52 | 86.7 |
| SWE-bench Verified | 71.10 | 71.9 |
| OSWorld-Verified (computer use) | 51.14 | 51.5 |
| OCRBench v2, English | 68.12 | 68.3 |
Newer models in the same table beat it on most tests. Terminal-Bench 2.1 shows the biggest gap:
| Model | Terminal-Bench 2.1 |
|---|---|
| Qwen3.8-Flash-Next | 85.80 |
| GLM-5.3-Flash | 82.70 |
| Nemotron 3.5 Super VL | 60.67 |
GLM-5.3-Flash is not ahead everywhere: Super VL beats it on OCRBench v2 and OSWorld-Verified. All figures are NVIDIA's own.
Where to run it
- Hugging Face: open weights for Super VL (BF16 and NVFP4) and Bolt.
- OpenRouter: listed at $0.15 per 1M input tokens and $0.50 per 1M output tokens. Some providers there charge less, from $0.10 and $0.40.
- ModelScope: in the NVIDIA community space.
- NVIDIA NIM: as a microservice on build.nvidia.com.
What to do next
- For document-heavy agents (PDFs, tables, screenshots), try Super VL on OpenRouter before you self-host.
- For a local setup on one 24 GB GPU, start with Bolt.
- Check the model cards for the reasoning settings: reasoning and the answer share the same
max_tokensbudget.
Sources
Did this land? Leave a reaction.
Discussion
0 commentsMembers can join the discussion
Become a member to comment and reply to other members.
Become a MemberNo comments yet
Be the first to start the discussion.