How to run NVIDIA's AI on your computer, with no internet
Nemotron 3.5 Bolt is an NVIDIA model that is free to download and runs on your computer offline.
Steps
- Download LM Studio, which is free: lmstudio.ai/download.
- Open it, search for "Nemotron 3.5 Bolt" and download the one from NVIDIA. This is the direct link: lmstudio.ai/nvidia/nemotron-3.5-bolt-e4b.
- Load it in the chat, turn off the wifi and ask it anything.
What Nemotron 3.5 Bolt is
NVIDIA released it on October 23, 2026. It starts from Gemma 4 E4B, Google's small model, and NVIDIA trained it to hold a conversation and use tools on the device itself: voice assistants, video game characters and robots. In NVIDIA's tests it scores about 40% higher on average than Nemotron 3 Nano 4B, its previous model of that size.
Specs
| Spec | Value |
|---|---|
| Released | October 23, 2026 |
| Made by | NVIDIA, starting from Google's Gemma 4 E4B |
| Download | 6.18 GB (GGUF format, Q4_K_M version) |
| Parameters | 7.5 billion |
| Memory it uses | Between 3.1 and 5 GB, depending on the machine |
| Context | 64,000 tokens |
| Languages | 35, Spanish among them |
| License | OpenMDW-1.1 |
| Downloads | 172,000 in the last month on Hugging Face, as of November 19, 2026 |
How fast it runs
| Machine | Tokens per second | Memory used |
|---|---|---|
| RTX 5070 (12 GB) | 133 | 3.5 GB |
| RTX Spark (32 GB) | 67 | 3.1 GB |
| Jetson AGX Orin (32 GB) | 24 | 4.2 GB |
| Jetson Orin Nano (8 GB) | 15.5 | 4.2 GB |
These are NVIDIA's measurements with the Q4_K_M version on llama.cpp, with a 3,000-token input and a 50-token answer. The data comes from research.nvidia.com and the model card on Hugging Face.