Nvidia released open-weight Nemotron 3.5 models - a 120B Super VL for enterprise agents and lighter Bolt variants that run locally on AI PCs with RTX Spark chips. For a small business this means private automation pipelines you can run on your own machine without paying per API call.
The TechCrunch report says Super VL is already in use at EY, Ramp, Salesforce and Perplexity, and Nvidia claims it beats Anthropic Opus 5 and GPT 6.1 on tool-use benchmarks when post-trained for agentic jobs - document intelligence, financial accounting, insurance claims. Bolt 74B-A9B 24GB targets local agents with multi-step reasoning and tool calling; the smaller Bolt E4B is for chat.
What this changes for small-business automation
Benchmarks are marketing until the model touches real client data. But a 24GB-friendly local model that handles tool calls shifts the math for document-heavy work - contracts, receipts, tender PDFs - where sending files to OpenAI or Anthropic is either expensive at scale or blocked by compliance. For a workflow like a legal archive audit pipeline I built that must stay inside the client's perimeter, a post-trained Bolt on one workstation starts to compete with cloud frontier calls on cost. The catch: post-training, evals and ops are now on you, not the vendor.
Why it matters: Open local models stopped being toys - for a shop that cares about data residency or API bills, running your own 74B agent on a workstation is a realistic option this quarter.
Source: TechCrunch