Nvidia's Nemotron 3.5 Super VL Scores 26 on Artificial Analysis, Plus a 2-Bit Bolt Model That Fits in 24GB. How They Compare

Nvidia now has an open model for every tier it sells hardware into, from data centres to laptops. The useful question is how the new models score when someone other than Nvidia runs them.

What Nvidia released

• Nemotron 3.5 Super VL: 120B total parameters, 12B active, 1M-token context, image and text in, text out. Available now on Hugging Face, OpenRouter, ModelScope and as an NVIDIA NIM microservice, with open weights, datasets and training recipes.
• Nemotron 3.5 Bolt 74B-A9B 24GB: 74B total, 9B active, 256k context. Uses native 2-bit quantisation to fit in 24GB of memory. On Hugging Face now, with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
• Nemotron 3.5 Bolt E4B 4GB: a small model for voice and conversational agents on GeForce RTX, Jetson and Spark devices. Coming 23 October.

How they score, measured by Artificial Analysis

Two results back up Nvidia's pitch. Super VL doubles its predecessor's score while using about 60% fewer output tokens, and at about 630 tokens per second it is the fastest model scoring 20 or more that Artificial Analysis has measured at that prompt length. Bolt comes within 1.5 points of Nemotron 3 Ultra with about a seventh of its total parameters, and beats it on AutomationBench-AA, but trails on long-context reasoning and Humanity's Last Exam.

The caveats

• The "leading US open-weight model" label is narrow. Artificial Analysis had Super VL at 26.5 against Inkling Small at 25.7. A day later, on 16 October, it published a score of 30 for Reflection's Beam, which it ranks second among US open-weight models, so Super VL no longer leads that group.
• Nvidia's claim that Super VL finishes tasks faster than Opus 5 and OpenAI's GPT 6.1 models applies after post-training on specific agentic tasks, as TechCrunch notes. It is not a general comparison.
• Bolt's token appetite: at 82k output tokens per task it uses 1.8 times as many as Super VL, which offsets some of its speed on local hardware.

Who is using it

Nvidia names Atlassian (search across Jira and Confluence), CrowdStrike (security), Dassault Systemes, Perplexity, Salesforce and Synopsys as platform companies integrating Super VL, plus organisations such as Ramp, EY and Hebbia post-training it through partners. These are Nvidia's own customer claims.

What this means if you run agents

• If you need an open model you can host and fine-tune on your own infrastructure, Super VL is now a fast, efficient option, but the strongest open weights still come from China and the UAE.
• If you want an agent that runs on a laptop with 24GB of memory, Bolt 74B-A9B is the model to try, with support in the usual local runners from day one.
• Either way, test on your own tasks. The headline scores are general benchmarks, and Nvidia's strongest claims depend on post-training.