NVIDIA unveils nemotron 3.5 bolt e4b 4gb: a compact language model for local AI applications

NVIDIA has launched Nemotron 3.5 Bolt E4B 4GB, an open-source small language model designed for local conversational agents. The model, trained using reinforcement learning from Google’s Gemma 4 E4B, is optimized for voice assistants, game NPCs, and robot interfaces on RTX, Jetson, and Spark hardware. It runs efficiently within 4GB of VRAM, allowing space for speech and vision models. The model’s deployment includes quantization to Q4_K_M with importance-matrix calibration, achieving 272.6 tokens/s on an RTX 5070. NVIDIA reports Bolt outperforms Gemma 4 E4B in speed, with 1.21x faster performance. It also leads in 15 of 21 benchmarks in reasoning mode, scoring 40% higher than Nemotron 3 Nano 4B. Scale Labs ranked Bolt first in MultiChallenge at 56.03%, though it trails Qwen3.5-4B in several benchmarks. Bolt is available on Hugging Face, Ollama, LM Studio, and NVIDIA’s NIM platform. NVIDIA emphasizes its use in local AI applications, including the Reachy Mini robot demo, where it handles speech replies and tool calls alongside Cosmos 3 Edge for vision.