NVIDIA Releases Nemotron 3.5 Bolt E4B 4GB: An Open Small Language Model for Local Voice Agents, Game Characters and Robots

NVIDIA has put out Nemotron 3.5 Bolt E4B 4GB, a small language model aimed at local conversational agents. The target use cases are pretty specific: voice assistants, game NPCs, and robot interfaces running on RTX, Jetson, and Spark hardware.

The pitch is memory, not scale-for-scale’s-sake. NVIDIA says the model needs about 4GB of VRAM for agent behavior, while still leaving headroom for speech and vision components. In its tests, the Q4_K_M build used between 3.1 and 4.3GB of VRAM on RTX hardware.

Under the hood, Bolt is a post-trained version of Google’s Gemma 4 E4B, and NVIDIA says it used reinforcement learning only. It skipped supervised fine-tuning entirely. The company says that choice preserved conversational quality better, even though SFT helped with tool calling. To make the model more useful for agents, NVIDIA added multi-turn tool use, a conversational_mode flag for shorter speech-friendly replies, configurable reasoning, and a runtime thinking budget to keep latency in check.

The benchmark picture is mixed, which is usually the honest version of these launches. NVIDIA says Bolt led 15 of 21 benchmarks in its own harness and hit 84.8 on IHEval, above Gemma 4 E4B and Qwen3.5-4B in that test. But it also trails badly on Tau2-Bench, where Qwen3.5-4B scored 75.1 and Bolt scored 55.1. It also loses on RULER 64K, LongBench v2, BFCL-v3, and EQ-Bench v3.

For deployment, NVIDIA quantized the BF16 checkpoint to Q4_K_M and says the release includes a DFlash drafter for speculative decoding. On an RTX 5070, Bolt with DFlash reached 272.6 tokens per second, faster than Gemma 4 E4B with its MTP drafter at 225.9 tokens per second. The model is available through Hugging Face, Ollama, LM Studio, Unsloth, Jetson AI Lab, NVIDIA NIM, and the In-Game Inferencing SDK.