NVIDIA releases Nemotron 3.5 Bolt E4B 4GB for local assistants and robots

Many accounts welcomed NVIDIA's Nemotron-3.5-Bolt-E4B for running locally on low-RAM edge devices like phones and robots, while some replies questioned benchmark credibility and asked for greater intelligence.

Based on 43 sentiment-bearing replies from 39 accounts across 2 conversations.

In NVIDIA's Reachy Mini demonstration, Bolt powers spoken responses and calls the robot's animation scripts. A separate model, Cosmos 3 Edge, supplies visual understanding. NVIDIA says both models run locally on a Jetson Orin Nano 8GB.

According to NVIDIA's model card, Bolt was further post-trained from Gemma-4-E4B-it, has 8 billion total parameters (4 billion effective) and accepts and produces text. NVIDIA lists multi-turn tool use, a mode for concise spoken replies, reasoning that can be switched on or off, and a context window of up to 64,000 tokens.

What the 4GB figure covers

NVIDIA says Bolt is designed to fit in as little as about 4GB in its target workloads. Its test table reports peak VRAM from 3.1GB to 5.0GB across listed RTX and Jetson configurations; results vary with hardware, software and configuration. The Hugging Face listing displays 6.18GB for NVIDIA's Q4_K_M model variant.

On its selected benchmarks, NVIDIA reports roughly 40% higher average scores than its earlier Nano 4B model using BF16 results. The company cautions that benchmarks only partly capture how an agent behaves in an application. Its model card says NVIDIA measured accuracy under NeMo Gym and that those figures may differ from vendors' reported results.

NVIDIA says the model is available on Hugging Face, along with weights, data and recipes for customization.

Many accounts welcomed NVIDIA's Nemotron-3.5-Bolt-E4B for running locally on low-RAM edge devices like phones and robots, while some replies questioned benchmark credibility and asked for greater intelligence.

Based on 43 sentiment-bearing replies from 39 accounts across 2 conversations.