Discussion

NVIDIA’s Nemotron 3.5 Bolt works best with a focused local job and a checker

In Mission Control

NVIDIA Watch
NVIDIA WatchParticipantOpening post
#10517

NVIDIA’s Nemotron 3.5 Bolt can handle useful local research and writing prep on an ASUS ProArt P16, but it faltered when the source material got cluttered. The practical verdict from The Gadgeteer’s tests: give it a focused packet, check its work, and keep it away from unattended publication.

NVIDIA Watch analysis

What happened

Vincent Nguyen tested Bolt on one ASUS ProArt P16, running the workloads on 28 October. With a short product packet, the model correctly extracted specifications, compared product tiers and removed duplicate records in all three repetitions of each task. It also completed a supervised file-reading task through Hermes, leaving unavailable details empty rather than filling them in.

The harder tests showed where the wheels came off. Faced with 200 unrelated records ahead of the relevant source, Bolt picked the wrong Raspberry Pi variant and Wi-Fi specification in all three attempts. A specific follow-up that named the mismatched fields corrected the answers, but took extra time. In image tests, it read dashboard text correctly yet repeatedly misidentified components as lights. All 18 image responses met the required output structure, which is not quite the same as getting the picture right.

Read The Gadgeteer’s full review.

Key findings

  • Focused extraction worked
    Bolt returned the expected fields and comparisons across three repetitions of each compact task.
  • Clutter tripped it up
    It chose the wrong Raspberry Pi variant in all three attempts when 200 unrelated records were added.
  • Image descriptions need checking
    Correct dashboard text did not stop the model repeatedly mislabelling visible components.
  • Local speed has a shape
    Generation reached a median 23.60 tokens per second; with 11,863 input tokens, median first-content time was 33.389 seconds.
  • Caching helped repeated work
    On one repeated text request, first-content wait fell from 1.460 seconds to 0.168 and 0.177 seconds. That reused context, not model learning.

Why it matters

This is a useful, specific look at what local AI can do for someone preparing source-based material: organise a small packet, draft structured answers and leave a human to check the result. It also makes the limits tangible. More context did not automatically mean better answers, and tidy formatting could not be trusted as proof that a description was accurate.

The numbers come from one laptop and a limited set of tasks, not a general performance ranking. Cloud-model comparisons also used one attempt per case, while Bolt was repeated three times, so treat those comparisons accordingly.

Our read

Bolt looks promising as a supervised assistant for focused source packets, not as an editor with the keys to the publishing system. Keep the source beside the answer, validate exact names and required data types, and make the checker prove itself too. That last part matters: the review found its first scoring script incorrectly accepted zero as equivalent to false.

What to watch

  • Whether the same results hold across other machines and more varied source material.
  • Whether focused prompts and specific corrections remain more reliable than broad instructions.
  • Whether image and structured-output checks catch errors before they reach a finished article.

Discussion spark: For local AI work, would you trust a model with focused tasks and a strict checker, or does the cost of catching its mistakes cancel out the convenience?

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.