Santa Clara-based Nvidia released Nemotron 3.5 Bolt E4B on Oct. 23, a small open model for conversational agents, device assistants and game characters that run locally.

Why it matters: A capable agent that fits in about 4 GB lets developers build private, offline assistants without per-call cloud costs.

What it does

  • Built on Google's Gemma 4 E4B with Nvidia's own post-training.
  • Supports multi-turn tool use, switchable reasoning and a conversational mode for short, speech-ready answers.
  • 64K-token context and 35 languages, per ToolSift's summary of Nvidia Research.
  • Deployment targets include RTX, Jetson and Spark systems.

Nvidia's claims

  • About 40% higher average than Nemotron 3 Nano 4B on conversational-agent benchmarks.
  • Needed about 3.5 GB of graphics memory on an RTX 5070 in Nvidia's tests.
  • Still trails Qwen 3.5 4B on some tests. These are vendor results, not independent benchmarks.
  • MUVO notes the BF16 release is mainly for customization, so the 4 GB figure depends on the variant.