[影片] 在 16 GB MacBook Pro 上測試 Nemotron 3.5 Bolt E4B 4GB

NVIDIA released Nemotron 3.5 Bolt E4B on October 23, 2026. It’s the smallest model in the new Bolt line — 8 billion total parameters with 4 billion effective, built by taking Google’s Gemma 4 E4B and running additional post-training on top of it. The focus is conversational agents, multi-turn tool calling, and lighter interactive or device-style use cases rather than heavy reasoning or pure coding. It supports a 64K context window, a thinking toggle, a conversational mode, and roughly 35 languages. NVIDIA also highlights that it produces around 80% less output tokens than similar models.

I tested the 4-bit version (available on Hugging Face) on my 16GB MacBook Pro, the same setup I use for other models in this series. On-disk size is between 4–6GB, depending on the exact quantization. The “4GB” in the name refers more to the target runtime memory envelope NVIDIA measured on certain hardware than to the download size itself.

The first check was a short open question: “What is a dolphin?” The model produced a solid, usable answer. It does some thinking before generating, and the tokens don’t appear instantly, but the total time stayed reasonable. For anything that doesn’t require maximum speed, it works reasonnably well.

The second check was a small coding task — generate a utility that counts words in files inside a folder. It chose to write a shell script rather than Python and delivered something that actually works. That was interesting; the model made a practical choice on its own.

Overall, Bolt E4B is usable on a 16GB MacBook Pro for day-to-day style work. It’s not the fastest model I’ve run in this memory class, but the answers are good enough and the token efficiency helps. Larger Bolt variants (the 74B class and above) exist for data-center or multi-machine home setups; those are outside what this series covers.

If you know other models that comfortably fit in 16GB and you’d like them tested the same way, leave a comment. If you like this kind of content, then like, share and subscribe to my YouTube channel!