NVIDIA releases 4GB Nemotron 3.5 Bolt model for robots and local agents, tested on Jetson Thor

NVIDIA said in the model card it published on Hugging Face and build.nvidia.com on Oct 23 that Nemotron 3.5 Bolt E4B 4GB, a small open language model it post-trained from Google's Gemma 4 E4B, is now available for local voice assistants, interactive characters and embedded interfaces, with a demo of Hugging Face's Reachy Mini robot holding a conversation.

The model has 8 billion parameters (4 billion effective), ships in Q4_K_M quantisation to fit about 4GB of memory, handles up to 64,000 tokens of context and was tested on Jetson Orin, Jetson Thor, DGX Spark and the GeForce RTX 5090; it is text in and text out, so it runs a robot's dialogue and tool calls, not its motors.

In NVIDIA's own quantised tests it scored 60.5 on the When2Call multi-turn benchmark against 42.9 for Alibaba's Qwen3.5 4B, but trailed Qwen on Tau2-Bench agent tasks (52.3 vs 61.7) and BFCL-v3 function calling (67.3 vs 70.6).

The licence (OpenMDW 1.1) allows commercial use. MEKAWIRE context: the release follows the GTC Berlin launches of Oct 21 and 22 (Isaac Lab 3.0, the FLORA data engine and new Cosmos tools) and fits the pattern our NVIDIA strategy explainer describes, free models that pull robot makers onto NVIDIA chips.