Nvidia Nemotron 3.5 Super VL and Bolt: open weights under OpenMDW, and what the Opus 5 comparison really shows

On October 15, 2026, Nvidia released Nemotron 3.5 Super VL, a 120-billion-parameter open-weight model built to run AI agents that read documents, charts, screens and video, alongside Nemotron 3.5 Bolt 74B-A9B, a compressed model for local machines. A third, Bolt E4B, a 4 GB model for voice and embedded agents, is due October 23. Super VL's and Bolt's weights come under OpenMDW-1.1, a license with no restrictions on how you use the model or its outputs. Nvidia's most eye-catching claim, that Super VL beats Claude Opus 5 and GPT-6.1 Sol, holds only for customized versions on specific tasks.

This story is based on Nvidia's blog and developer posts, the Super VL and Bolt 74B-A9B model cards on Hugging Face, the OpenMDW-1.1 license text, Artificial Analysis's published results, and reporting from VentureBeat, checked October 16, 2026.

The short version

• Super VL: 120B total, 12B active parameters; a Mamba-2/Transformer hybrid MoE with 512 routed experts per MoE layer; text, image and video input; up to a 1M-token context.
• Bolt 74B-A9B: 74B total, 9B active, a compressed version of Super VL natively quantized to 2 bits to fit in 24 GB; up to a 256K-token context. Available on Hugging Face.
• Bolt E4B: a 4 GB model for GeForce RTX, Jetson and Spark systems, due October 23.
• License: OpenMDW-1.1 for Super VL and Bolt 74B-A9B, with no use restrictions. Nvidia also published training data, recipes and RL environments.
• Claims: vendor-reported. The wins over Opus 5 and GPT-6.1 Sol come from post-trained versions on narrow tasks.

What did Nvidia release?

Super VL is a vision-language version of the Nemotron 3 Super line. Per the model card, its language backbone was initialized from Nemotron 3 Super Base, and Nvidia's developer post says it was distilled from Nemotron 3 Ultra. It pairs Mamba-2 layers, which process long sequences more cheaply than attention, with Transformer attention layers and a C-RADIO vision encoder.

Last verified: October 16, 2026 (Nvidia blog and developer posts, Super VL and Bolt 74B-A9B model cards)

Nvidia says Super VL was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with about 3.8 million samples and reinforcement learning. The release includes BF16 and NVFP4 checkpoints, a base model, datasets on Hugging Face, and RL environments in NeMo Gym. The model card's deployment examples use 1M-token contexts, but some hosts and recipes use shorter windows, such as 131,072 or 262,144 tokens.

Bolt 74B-A9B targets local coding and research agents on RTX and DGX Spark machines, with support in llama.cpp, Ollama, LM Studio, Unsloth and vLLM, per Nvidia's blog. Its model card describes it as a compressed version of Super VL that keeps the same text, image and video input. Nvidia's blog calls Bolt "the #1 open model in its class on PinchBench"; in the comparison table on Bolt's own model card, it scores 87.82 on PinchBench, just behind Qwen3.6-35B-A3B at 88.07, and leads on BrowseComp with 66.75. Both are Nvidia-run figures. If you're sizing a model for your own hardware, our guide to choosing a small model for your laptop explains the memory math.

What does OpenMDW allow?

OpenMDW-1.1, published on a Linux Foundation-copyrighted site, grants permission "to deal in the Model Materials without restriction," covering copyright, patent, database and trade-secret rights. Its conditions are short:

• Keep a copy of the license and the copyright notices when you redistribute.
• No use restrictions or acceptable-use policy, and none on outputs.
• Your rights end if you sue claiming the materials infringe a patent or copyright, unless you're responding to a suit brought against you.

That makes it more permissive than the custom licenses in our open-weight license comparison: no user cap, no revenue gate, no branding rule. It also stands apart from two other October releases: Liquid AI's LFM2.5, which excludes companies with $10 million or more in revenue, and Meta's Muse Spark Open, which requires branding and gates large hosting businesses. Because the training data and recipes are published too, the release goes further toward the Open Source Initiative's definition than weights-only launches do.

Does Super VL really beat Claude Opus 5 and GPT-6.1 Sol?

Not in general. Nvidia published a chart of five specialized tasks (document intelligence, security attack resilience, financial receipt extraction, scientific experimentation and medical insurance claims) comparing Super VL before and after task-specific post-training. VentureBeat's Carl Franzen read two named comparisons from it:

• Medical claims: about 28% for the base model, about 72% after post-training, against about 38% for Claude Opus 5.
• Security attack resilience: about 35% to about 76%, against about 66% for GPT-6.1 Sol.

These are vendor-reported. Nvidia hasn't published task definitions, sample sizes or the testing protocol, and, as VentureBeat put it, "these are different customized versions on different specialized tasks." The comparison shows that a small model tuned on your data can beat a general frontier model on that job. It doesn't show Super VL is a stronger general model.

On broader measures:

• Artificial Analysis Intelligence Index: On October 15, Artificial Analysis scored Super VL 26 (26.5 unrounded) on its Intelligence Index v4.3.2, a composite of 10 evaluations it runs itself, and Bolt 74B-A9B 21. It called Super VL the leading US open-weight model at that point, ahead of Thinking Machines' Inkling Small (25.7) and Nvidia's Nemotron 3 Ultra (23); Meta's Muse Spark Open, released the same day, has since been listed at 39. Chinese open models score higher, for example GLM-5.3-Flash at 42. None of the index's 10 evaluations is an image test, which is Super VL's focus.
• Speed: Artificial Analysis reported about 630 output tokens per second with 100,000-token prompts on a pre-release DeepInfra endpoint serving the NVFP4 weights. Nvidia cites about 625 tokens per second on four B200 GPUs, against about 125 for Inkling Small, VentureBeat reported. That's generation speed, not end-to-end task time.
• Model card results (Nvidia-run): 71.1% on SWE-bench Verified, 89.5% on GPQA-Diamond and 80.8% on CharXiv reasoning for the BF16 checkpoint. Nvidia's card compares only its own BF16 and NVFP4 versions.

Customer results Nvidia cites, such as a 40x inference cost cut for Scale AI's work for Skydance, are also vendor-supplied, and Nvidia's own post is inconsistent on that project's latency gain (5x in the body, 10x in the summary).

Why does it matter?

Nvidia ships its large open models together with training data and recipes, and Super VL fits the hardware story it's telling, from the data center down to DGX Spark desktops and Vera Rubin racks. Nvidia notes the models are free, while production support through Nvidia AI Enterprise costs $4,500 per GPU per year, according to VentureBeat. For teams that need to post-train a document or vision agent on their own data and ship it without license strings, Super VL is a serious option.

What's confirmed, reported and uncertain

• Confirmed (Nvidia, Hugging Face): October 15 release; Super VL's 120B/12B hybrid MoE design, 1M context, inputs and OpenMDW-1.1 license; Bolt 74B-A9B's 2-bit, 24 GB design, 256K context and OpenMDW-1.1 license; Bolt E4B due October 23; released datasets and recipes.
• Third-party: the Intelligence Index scores (26 for Super VL, 21 for Bolt) and the ~630 tokens-per-second speed measurement (Artificial Analysis).
• Vendor-reported: the task-specific wins over Opus 5 and GPT-6.1 Sol, Nvidia's speed claims, PinchBench and BrowseComp results for Bolt, model card scores and customer results.
• Reported: the $4,500 support price (VentureBeat).
• Uncertain: Bolt E4B's parameter count, context and license; independent results on Super VL's vision and agent tasks.