NVIDIA releases Nemotron 3.5 Bolt E4B small language model

NVIDIA has released Nemotron 3.5 Bolt E4B, the smallest model in its Bolt series, designed for local conversational agents such as voice assistants, game NPCs, and robotic interfaces. The model is derived from Google’s Gemma 4 E4B-it base, with NVIDIA applying reinforcement learning via NeMo RL and NeMo Gym (GRPO) to enhance its ability to follow instructions and call tools accurately during multi-turn conversations.

Targeting low-memory environments, the model aims for approximately 4GB of memory usage depending on the workload and quantization. For instance, tests on an RTX 5070 showed memory usage between 3.5GB and 4.3GB. The model supports a context length of 64K tokens and includes a reasoning option that can be toggled to manage latency.

Performance benchmarks indicate improvements over previous Nano 4B models, with NVIDIA reporting up to 40% higher scores in certain conversational agent tests. The model is capable of executing API functions through tool calling, though actual data access is determined by the specific application integration.