Nemotron 3.5 Bolt E4B 4GB (Reasoning)

Nemotron 3.5 Bolt E4B 4GB (Reasoning) is an open small language model released by NVIDIA, built on top of Google's Gemma 4 E4B architecture with specialized post-training for conversational agents and edge devices. Featuring an 8-billion parameter total size with 4 billion active parameters (E4B), the model is optimized for high intelligence-per-gigabyte efficiency, fitting standard local runtime requirements into roughly 4GB of VRAM. It is engineered specifically to balance low memory overhead with responsive processing capabilities across resource-constrained environments.

Capabilities and Architecture

The model incorporates configurable reasoning, allowing users to enable explicit thinking processes for complex analytical requests or disable them to prioritize minimal latency. It supports multi-turn tool use for executing application programming interfaces across extended dialogues, alongside a specialized conversational mode tailored for concise, speech-ready outputs. Additionally, Nemotron 3.5 Bolt E4B includes support for a 64k context window and accommodates multilingual interactions across numerous languages. Speculative decoding support via companion drafters like DFlash helps further enhance token generation throughput on local hardware configurations.