Nvidia’s new Nemotron models split the agent race between servers and local PCs
Super VL and Bolt target different deployment sizes, making memory requirements and task reliability central to the comparison.
Nvidia released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B on October 15, presenting a family that spans server-based and local agents. A smaller Bolt E4B model is scheduled for October 23. Nvidia’s release announcement.
Super VL has 120 billion parameters and targets work involving text, images and structured documents. Nvidia says Bolt 74B-A9B uses native two-bit quantization to fit within 24GB of memory. The company is releasing weights, datasets and training recipes. Its comparative speed and accuracy statements are vendor claims, not GeekRadar benchmark results.
Deployment becomes part of the model choice
Our assessment is that this announcement makes the practical buying question more interesting than a single leaderboard position. An agent that reads a document, selects a tool and checks a result needs to be evaluated as a complete workflow. Memory footprint determines where it can run; successful completion determines whether running it is worthwhile.
For a local deployment, we would examine the room left for other applications and the cost of keeping useful context. For a shared server, we would ask how competing requests affect responsiveness. Neither answer follows directly from the headline parameter count.
Openness makes scrutiny possible
Access to model artifacts can make it easier to inspect, adapt and reproduce a system. The next useful step is to see how clearly a particular task can be repeated under documented settings.
A convincing comparison would identify the model version, hardware, tool permissions and pass criteria. It would also show the failed attempts. Otherwise, two demonstrations that look similar may be testing very different amounts of assistance.
Our DeepSeek-R1 retrospective examined why inspectable training work matters. The same editorial question applies here: what can another team independently check?
What to watch next
Readers can follow our AI section and on-device AI topic. The useful follow-up will be evidence about finished tasks on representative systems. Until then, this release supplies new deployment options and specific claims to test, rather than a universal answer about which agent model is best.
