NVIDIA Releases Nemotron 3.5 Bolt E4B 4GB for Local Voice Agents
NVIDIA released Nemotron 3.5 Bolt E4B 4GB, an open small language model for local voice agents, game characters and robots, with about 4GB VRAM use and RL-only post-training from Gemma 4 E4B.
The model has 8 billion total parameters and 4 billion effective parameters, with a 64K context window for conversation history, tool definitions and application state. NVIDIA said the Q4_K_M build used 3.1 to 4.3 GB of VRAM on RTX in its tests. The launch lists three target workloads: device and voice assistants, interactive characters, and embedded or industrial agents. In a research demonstration, a Reachy Mini robot ran on a Jetson Orin Nano 8GB, with Cosmos 3 Edge handling vision while Bolt handled speech replies and tool calls.
NVIDIA post-trained Bolt from Google’s Gemma 4 E4B using reinforcement learning only, according to the report. The research team skipped supervised fine-tuning, saying SFT improved tool calling but hurt Gemma’s conversational quality and verbosity. It used RLVR and RLHF through NeMo-RL and NeMo-Gym. Every training sample ran in both reasoning and non-reasoning modes, and each batch kept a 2:1 ratio of non-reasoning to reasoning samples. Length penalties in the GenRM environment stopped answers from growing during training. NVIDIA said the datasets and full recipe are public.
The launch post lists multi-turn tool use with guidance for asynchronous tool calls, a conversational_mode flag for short speech-ready replies, configurable reasoning with a runtime thinking budget to cap latency, and the 64K context. NVIDIA also calls Bolt the third-generation small model for its ACE game characters.
NVIDIA measured 21 benchmarks across five capability areas in its NeMo-Gym harness, all in BF16. Bolt led 15 of 21 rows in reasoning mode. Its best reported result was IHEval 84.8, versus 67.0 for Gemma 4 E4B and 60.6 for Qwen3.5-4B. Its weakest listed result was Tau2-Bench 55.1, about 20 points behind Qwen3.5-4B at 75.1. NVIDIA also reported roughly 40% higher average scores than Nemotron 3 Nano 4B and claimed up to 80% fewer output tokens in reasoning-enabled comparisons. Separately, Scale Labs ranked Bolt first among small models on MultiChallenge at 56.03%, according to an X post cited by MarkTechPost.
For deployment, NVIDIA quantized the BF16 checkpoint to Q4_K_M with importance-matrix calibration using about 1 million tokens of RL rollouts. The release includes a DFlash drafter for speculative decoding. On an RTX 5070, Bolt with DFlash reached 272.6 tokens per second, while Gemma 4 E4B with its MTP drafter reached 225.9 tokens per second, making Bolt 1.21 times faster in that NVIDIA test.