NVIDIA Releases Nemotron 3.5 Bolt E4B: Local Agents on 4GB of VRAM

What happened

Bolt E4B is the final piece of the Nemotron 3.5 family NVIDIA announced Oct 15: Super VL (120 billion parameters) and Bolt 74B-A9B (24GB) were already available, with E4B scheduled for Oct 23. On Oct 23, the NVIDIA AI account posted a demo: a Reachy Mini robot chatting and moving its head and antennas on an 8GB Jetson Orin Nano, with Bolt handling conversation and Cosmos handling vision, both running locally.

Structurally, Bolt E4B has 8 billion total parameters and 4 billion effective parameters, built on Google's open Gemma 4 E4B and post-trained by NVIDIA with reinforcement learning. Weights are a free download from Hugging Face under the OpenMDW-1.1 license, which permits commercial use, and the model runs with llama.cpp or Ollama.

The name needs one clarification: the 4GB in the name refers to NVIDIA's VRAM target, while the actual GGUF download is about 6.18GB and running it needs at least 4 to 5GB of VRAM. NVIDIA positions E4B for fast conversational agents, as distinct from the desktop-class Bolt 74B or the data-center-class Super VL.

Key facts

1. On Oct 15 NVIDIA released Nemotron 3.5 Super VL and two Bolt models; Super VL and Bolt 74B-A9B were available then, Bolt E4B was scheduled for Oct 23 and arrived on schedule.
2. Bolt E4B has 8 billion total parameters and 4 billion effective parameters, with E standing for effective; it starts from Google's Gemma 4 E4B and was fine-tuned by NVIDIA with reinforcement learning.
3. Target uses are local voice assistants, interactive characters and robots that converse and use tools; in the demo, Reachy Mini ran locally on an 8GB Jetson Orin Nano, with Bolt handling conversation and Cosmos handling vision.
4. Weights are a free Hugging Face download under OpenMDW-1.1, which permits commercial use; it runs via llama.cpp or Ollama, with a GGUF file of about 6.18GB.
5. Family roles: Lightning targets high throughput, Ultra targets the hardest reasoning tasks, Super VL targets always-on agents, and Bolt targets edge and local devices.