NVIDIAAnalysisNVIDIA releases open Nemotron 3.5 Bolt models for local multi-turn agentic workflows
What happened.NVIDIA published weights and architecture specifications for Nemotron 3.5 Bolt, a family of open edge language models engineered for local multi-turn agent execution and tool use.
Why it matters.Cloud-tethered agent loops introduce significant network latency and high per-token operating costs for persistent background tasks. Delivering quantized hybrid architectures that fit within 24 gigabytes of VRAM allows developers to run autonomous agents and tool-calling loops directly on local workstations. Hardware-level efficiency accelerates adoption in robotics, industrial automation, and private code generation.
The technical specifications:
- Model configurations: The release includes the 4-gigabyte Nemotron 3.5 Bolt E4B built on Gemma 4 architecture and the 24-gigabyte Bolt 74B-A9B hybrid Mamba-Transformer Mixture-of-Experts model.
- Edge optimization: Bolt 74B-A9B uses 2-bit Quantization-Aware Distillation to compress the 120-billion parameter Super VL model into consumer RTX and DGX Spark memory envelopes while retaining high decoding throughput.
- Agentic primitives: Models feature native 64,000-token context windows with configurable thinking depth, asynchronous tool-calling instructions, and a low-latency conversational mode for voice assistants.
STORYLINE PROGRESSION & CONTEXT
This report represents a verified evolution of prior newsroom intelligence: