Nemotron 3 Nano coding test with Nanocoder

Nvidia's Nemotron 3 Nano gets a coding test in Nanocoder: write a project brief, then build a browser particle simulator with vanilla HTML, CSS and JavaScript. The requested controls include dragging to add particles and choosing particle types, with gravity and collisions governing their movement.

The speaker uses Ollama Cloud for this test rather than running inference on his Mac. He says his 24 GB unified memory machine is close to its performance limit while recording at 1080p. For viewers considering this local LLM, he suggests roughly 24 GB of VRAM or unified memory for a 4-bit quantized version. That is his hardware guidance, not a measured local performance result.

The brief earns a positive assessment, but implementation requires repeated intervention. The model issues irrelevant commands, breaks its code and struggles to debug errors. After clearing context and restarting, the speaker narrows the project to one HTML file. He eventually gets moving particles with gravity, adjustable size and color, and click interaction. Bugs remain; changing the emission rate appears to require stopping and restarting the simulation.

The speaker reports spending over an hour on the task and prefers his earlier experience with Ministral 8B as a coding assistant. He does not recommend Nemotron as his daily driver, while acknowledging that Nanocoder compatibility or the served quantization could have affected the result.