Nemotron 3.5 Makes Nvidia’s Open A.I. Model Twice as Smart
Trending Topics Jakob Steinschaden ● Covered by 2 sources
Nvidia’s new open AI model is out, and it’s twice as smart as the last one. It also spits out far less text, which could make it cheaper and faster to use.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nvidia isn’t just selling the picks and shovels anymore. Last week, it put out Nemotron 3.5 Super VL, an open model that Artificial Analysis says is twice as intelligent as its predecessor while producing about 60% less text. It also launched Nemotron 3.5 Bolt, a family built to run on laptops and other local machines.
Super VL is a hefty system: 120 billion parameters in total, with 12 billion active on each request. It handles text, images, web pages and PDFs, and Nvidia says it is particularly strong at tables, charts and scanned documents. The model can take in a million tokens of context, which is the kind of number that sounds abstract until you remember it’s aimed at corporate agents chewing through long workflows, software tools and business files.
Nvidia is also being more open than many rivals. It released the weights under the permissive OpenMDW-1.1 license and says it is publishing training datasets and recipes too. The model was trained on about 26 trillion tokens, then post-trained with roughly 3.8 million examples plus reinforcement learning. It’s already available on Hugging Face, through OpenRouter and as an Nvidia microservice. Artificial Analysis pegs API pricing at about €0.21 per million input tokens and €0.51 per million output tokens.
On the benchmark side, Super VL scores 26 points on the Artificial Analysis Intelligence Index, which bundles 10 tests across coding, agents, knowledge work and science. Its predecessor, Nemotron 3 Super, scored 13. Nvidia’s previous flagship, Nemotron 3 Ultra, scored 23. Super VL also turns out to be relatively lean at output: 46,000 tokens per test task on average, versus 114,000 for the older model. And it’s fast. DeepInfra measured about 630 tokens per second on very long 100,000-token inputs, with the first word arriving after 1.8 seconds. Artificial Analysis says that is the fastest it has seen for any model scoring at least 20 points.
Bolt is the local counterpoint. The first model in that family has 74 billion parameters, 9 billion active, and with 2-bit quantization it fits into 24 gigabytes of memory. That puts it on gaming PCs with RTX cards, Nvidia’s DGX Spark mini computer and the new RTX Spark laptops that went on sale last Friday. It supports Ollama, LM Studio and llama.cpp. Bolt scores 21 points, just behind Nemotron 3 Ultra, but it uses 82,000 tokens per task, almost twice as many as Super VL. Nvidia says a smaller Bolt E4B is coming Friday for voice assistants and robots, and that one needs just 4 gigabytes.
Nvidia is clearly trying to make Nemotron the base layer other companies build on. Salesforce has already used it for its Koa reasoning model, and Nvidia lists Atlassian, CrowdStrike, Perplexity, Synopsys and Dassault Systèmes as partners. Through Scale AI, Super VL was also adapted for a due diligence agent at EY. The company’s reported talks to buy the open-weight startup Reflection AI only make that strategy louder.
My take — AI-written commentary, not fact-checked reporting
This is the sane part of the AI race: smaller, faster, cheaper, and actually shippable. Nvidia knows open models win adoption when they are useful, not when they win a benchmark trophy and then melt a server rack. Europe keeps talking sovereignty; Nvidia is quietly selling the tooling that makes it real.
Read more about this at: Trending Topics
