NVIDIA NEMOTRON 3.5 · P.K. SHARMA ANALYSIS
Explore the trade-offs
Independent benchmark position
Artificial Analysis Intelligence Index · Higher is better
Selected comparators, not the full leaderboard. New releases in teal. Select a row for context.
Data: Artificial Analysis release evaluation · 15 October 2026 · Index v4.3.2. Model specifications: NVIDIA.
Same laptop, different configurations
Surface Laptop Ultra · RTX Spark · 128GB · llama.cpp
Common bar scale: 0–110 output tokens/s. Higher is better.
Source: NVIDIA Bolt research release, 15 October 2026. Output: AIPerf SPEED-Bench-Coding. Duration: AA-AgentPerf Local.
A 24GB model is only the first allocation
Editable planning exercise · Not a hardware compatibility test
Teal: reported packed model. Blue: assumed runtime. Gold: assumed drafter. Unfilled: arithmetic headroom.
24 + 12 + 2 = 38GB allocated
24GB packed model: NVIDIA Bolt card. All other allocations are editable assumptions. GB here is a nominal planning unit.
From Super VL to local Bolt
Published development sequence, simplified · Select each stage
Source: NVIDIA Bolt model card and research explanation. This is a simplified account, not an executable training recipe.
Motion and tab switching need JavaScript. The benchmark values above remain readable. The article provides specifications, results and limitations as ordinary text and tables.