Nvidia

Nemotron 3.5 Makes Nvidia’s Open A.I. Model Twice as Smart

co-produced by AI19. October 2026, 18:47

Nvidia Nemotron 3.5 key visual — Nemotron 3.5, Nvidia’s new family of open models. © Nvidia

Nvidia no longer just sells the shovels for the A.I. gold rush, it is now digging for itself: Last week, the chipmaker released Nemotron 3.5 Super VL, an open model that, according to the benchmarking service Artificial Analysis, is twice as intelligent as its predecessor while producing about 60 percent less text. Nvidia also introduced Nemotron 3.5 Bolt, a new family of models designed to run on laptops.

120 Billion Parameters and a One-Million-Token Context Window

Nemotron 3.5 Super VL has 120 billion parameters, 12 billion of which are active per request. Besides text, the model processes images, web pages and PDFs, and Nvidia says it is especially good at reading tables, charts and scanned documents. Its context window holds one million tokens. It is built for corporate A.I. agents that work through multistep processes, operate software tools and review business documents.

Nvidia discloses more than many rivals: Along with the weights, released under the permissive OpenMDW-1.1 license, the company is publishing training data sets and recipes. The model was trained on about 26 trillion tokens and then post-trained with roughly 3.8 million examples and reinforcement learning. It is available on Hugging Face, through OpenRouter and as an Nvidia microservice. According to Artificial Analysis, API access costs about €0.21 ($0.23) per million input tokens and €0.51 ($0.57) per million output tokens.

Faster and Leaner Than Its Predecessor

On the Artificial Analysis Intelligence Index, which combines 10 evaluations covering coding, agents, knowledge work and science, Super VL scores 26 points. Its predecessor, Nemotron 3 Super, scored 13, and Nvidia’s previous flagship, Nemotron 3 Ultra, scored 23. The new model generates an average of 46,000 output tokens per test task, compared with 114,000 for its predecessor.

Nvidia has also picked up speed. On a pre-release server run by the provider DeepInfra, Super VL reached about 630 tokens per second with very long inputs of 100,000 tokens. According to Artificial Analysis, that is the highest speed it has measured for any model scoring at least 20 points. The first word of a response arrives after 1.8 seconds. One reason is the 4-bit format NVFP4, which runs particularly efficiently on Nvidia chips. According to Artificial Analysis, it costs only 0.1 points of quality compared with the full weights.

Bolt: An Agent for Your Laptop

The second launch, Nemotron 3.5 Bolt, is built to run locally. The first model in the family has 74 billion parameters, 9 billion of them active, and thanks to 2-bit quantization it fits into 24 gigabytes of memory. That means it runs on gaming PCs with RTX graphics cards, on Nvidia’s DGX Spark mini computer and on the new RTX Spark laptops, which went on sale last Friday. Popular tools such as Ollama, LM Studio and llama.cpp are supported.

Bolt scores 21 points on the Intelligence Index, just behind Nemotron 3 Ultra, which at 550 billion parameters is more than seven times its size. According to Artificial Analysis, it ranks fourth among models with fewer than 100 billion parameters and first among U.S. models in that class. The catch: Bolt is chattier, using 82,000 tokens per task, almost twice as many as Super VL. On Friday, Nvidia plans to follow up with Bolt E4B, a mini version for voice assistants and robots that needs just 4 gigabytes.

Who Already Uses Nemotron

Nvidia is deliberately positioning Nemotron as a foundation for other companies’ models. Salesforce has already built its reasoning model Koa on Nemotron, and Nvidia now lists partners including Atlassian, CrowdStrike, Perplexity, the chip design specialist Synopsys and the French software group Dassault Systèmes. Through Scale AI, Super VL was also post-trained for a due diligence agent at EY. How serious Nvidia is about its own models is also evident from its talks about acquiring the open-weight start-up Reflection AI.

A Lead With a Caveat

Artificial Analysis calls Super VL the best open model from the U.S. Close behind is Thinking Machines’ Inkling Small with 25.7 points. That title only applies to permissively licensed models, though: Meta’s Muse Spark Open, released the same day, scores 39 points on Artificial Analysis, but with 816 billion parameters it is much larger and comes with its own license that carries restrictions.

Internationally, the gap remains wide. The best open model from China, MiMo-V2.6-Pro, reaches 46 points, and K2 Horizon from the United Arab Emirates scores 31. Europe’s great hope, Mistral, scored 38 points on Artificial Analysis with Large 4. Nvidia is placing a different bet: Instead of building the smartest open model, it wants to deliver the fastest and cheapest one for agents that run around the clock. More details are likely on Wednesday, when Nvidia gives its keynote at GTC in Berlin.