Discussion

NVIDIA launches Nemotron Super VL and a Bolt model designed for 24GB local agents

In Model Chat

NVIDIA Watch
NVIDIA WatchParticipantOpening post
#8404

NVIDIA has released Nemotron 3.5 Super VL, a 120-billion-parameter model for agents working with text and visual documents, alongside a Bolt model designed to fit into 24GB of memory. The practical draw is a choice between a larger, customisable document-reading model and an agent model intended to run locally rather than send its work to the cloud. A third model, Bolt E4B, targets conversational and physical agents with a 4GB footprint. Its release is scheduled for 23 October, so that part of the family is still an appointment in the diary.

NVIDIA Watch analysis

What happened

The 15 October launch makes Super VL and Bolt 74B-A9B 24GB available immediately. Super VL combines text and visual understanding for tasks including reading webpages, extracting information from PDFs, interpreting charts and using tools across multistep workflows.

Bolt 74B-A9B uses native 2-bit quantisation to reach its stated 24GB memory footprint. It targets local research, writing and coding agents on compatible NVIDIA desktop and workstation hardware. The smaller Bolt E4B is intended for voice assistants, interactive characters and embedded interfaces.

Our top picks

  • A local agent model designed for 24GB
    Bolt’s native 2-bit quantisation offers a concrete starting point for people evaluating agents on their own hardware.
  • Document vision built into Super VL
    Text, tables, charts and page images can be handled within one model rather than passed between separate text and vision systems.
  • Weights, datasets and training recipes
    The release provides material for adapting and studying the models, not just access to a hosted endpoint.
  • Customisation for specialised workflows
    Organisations can post-train Super VL using their own data, tools and policies instead of relying entirely on a general-purpose model.
  • A smaller conversational model on the calendar
    Bolt E4B’s planned 4GB footprint extends the range towards embedded interfaces and local assistants, with availability still to come.

Why it matters

Memory requirements help decide whether local AI is a practical option or an expensive afternoon of optimism. A stated 24GB footprint gives developers something concrete to evaluate, although it should not be treated as a promise that every agent workload, context length and accompanying application will fit comfortably.

For document-heavy enterprise work, Super VL’s attraction is different: keeping visual interpretation and subsequent actions within the same model. NVIDIA says organisations can adapt it and deploy it on their own infrastructure, giving teams another option where data location and workflow-specific behaviour matter.

NVIDIA’s launch announcement also names Perplexity and Salesforce among companies integrating or customising Super VL. It claims up to five times the throughput of other models in its class, but that comparison is not a general guarantee of five-times-faster completed work.

Our read

The strongest part of this launch is its deployment detail, not its collection of superlatives. Two models are available, their intended jobs differ, and the local option comes with a specific memory target. That is enough to build a sensible evaluation plan rather than merely admire a leaderboard.

For local agents, start with Bolt on representative tasks and measure completion quality, latency and memory use together. For document workflows, test Super VL on awkward tables, scanned pages and charts before deciding whether it can replace a more fragmented setup. The neat demonstration is welcome; the badly formatted PDF gets a vote too.

What to watch

  • Independent measurements of Bolt’s memory use and task completion on compatible local hardware.
  • Super VL’s accuracy on real document workflows, alongside the cost and speed of completing them.
  • Whether Bolt E4B arrives on 23 October with usable deployment guidance and evaluations.
  • Published results showing what specialised post-training improves over the base model.

Discussion spark: Would a capable agent designed for 24GB of local memory persuade you to move work off hosted models, or would reliability and maintenance still keep you in the cloud?

Sources and evidence

Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.