Nemotron 3.5 Bolt E4B 4GB went live at 15:41Z on October 23, according to NVIDIA AI's announcement. The company describes it as a small open model built for local assistants, characters, and robots, with tool use included as a core capability rather than an add-on. The release targets on-device deployment, meaning inference meant to run without a round trip to a remote server.
The announcement itself is the evidence, posted by NVIDIA AI and flagged as an opened, supported release with high discovery confidence from the source feed. That status indicates the model is live and documented, not a preview or a limited rollout. No benchmark scores, latency figures, or hardware requirements beyond the 4GB footprint appear in the announcement, so claims about speed or accuracy on specific chips remain unverified until NVIDIA or third parties publish them.
What the release demonstrates is availability and positioning: a 4GB model, under the E4B designation, aimed squarely at edge and embedded use cases where cloud inference is impractical. What it does not demonstrate is comparative performance against other small open models, nor does it specify which license governs commercial use, nor which robotics platforms or SDKs have confirmed compatibility. NVIDIA did not disclose training data composition or evaluation results in the announcement, a gap worth noting before drawing conclusions about capability.
For teams evaluating on-device assistants, the tool-use framing matters more than the parameter count. A 4GB model that can call external functions reliably changes what a robot or embedded assistant can do without network access, provided the tool-calling behavior holds up under real workloads rather than curated demos. Operators should check quantization options, context window limits, and whether the model ships with a reference integration for common robotics middleware, none of which the announcement addresses. Researchers comparing this release against other sub-5GB open models will want independent benchmarks before citing capability claims.
The gap between an announcement and a reproducible evaluation is where most small-model claims either hold up or quietly dissolve, and Nemotron 3.5 Bolt E4B has not yet crossed that gap in public view. Until third-party testing on representative robotics and offline-assistant tasks appears, the safest reading is that NVIDIA has shipped a model sized for the edge, with tool use as its stated differentiator, and left the performance case to be made later by whoever runs it first.