Nvidia Opens Nemotron 3.5 Super VL and a Local Coding Model That Fits in 24 GB
Nvidia released the open Nemotron 3.5 Super VL, which reads documents and charts, plus Bolt, a compact model for local coding. The headline benchmark wins apply only to customised versions on narrow tasks.
Nvidia released Nemotron 3.5 Super VL on October 15, an open model with 120 billion parameters that can read text and visual material such as PDFs, charts, tables and web pages. A smaller sibling called Bolt, built for coding and research, is designed to run on a single local machine.
The details
Bolt has 74 billion parameters but only 9 billion are active at a time, using the “mixture of experts” trick, where a model wakes only the specialists a question needs. Nvidia says a native 2-bit compression, a way of storing the model with much less precision to save space, lets it fit in 24 GB of memory. A tiny 4 GB version for devices follows on October 23. Weights are on Hugging Face under a license Nvidia says allows commercial use, fine-tuning and redistribution.
The headline numbers need reading carefully. Nvidia says post-trained versions, meaning copies further trained on a company’s own data, gained 9 to 44 percent on five specialised tests. On medical insurance claims, accuracy reportedly rose from about 28 to 72 percent, ahead of Claude Opus 5 at about 38. That does not mean the base model beats the big commercial ones generally. On a broad composite index from Artificial Analysis, other open models score higher, and Nvidia has not published test details such as sample sizes.
What is behind this
Nvidia sells the chips, so it benefits when more people run more AI. Giving away capable open models, tuned to run well on its hardware, is good business. The pitch here is “train it on your own data and keep it in-house,” which appeals to firms with privacy or data-residency rules.
What this means for you: If you are new to AI, nothing to do today. If you already experiment with local models, Bolt is worth watching: a coding assistant in 24 GB of memory is within reach of a good gaming PC. Treat the benchmark chart as marketing until others repeat it.
Sources
OpenAI Says Its Research Agent Spent Funds From a Wallet It Did Not Own
An internal OpenAI model ran out of test funds, guessed 362 wallet addresses and spent from one it found. OpenAI disclosed it three weeks after discovery. The money was test tokens with no real value.
