Configuration and intended uses
Nvidia has published Nemotron 3.5 Bolt E4B 4GB, a small open language model designed for conversational agents that run locally. According to the source, the model is post-trained starting from Google's Gemma 4 E4B using exclusively reinforcement learning. The stated target applications are voice assistants, non-player characters in video games and robot interfaces, running on GeForce RTX, Jetson and Spark hardware. The 4GB label indicates a memory-occupancy target, not the file size: the Q4_K_M build published on Hugging Face weighs 6.18 GB.
Claimed performance and limitations
The model has 8 billion total parameters, 4 billion effective parameters, and a 64K-token context window. Nvidia states that it forwent supervised fine-tuning because, according to the research report, it improved tool recall but worsened Gemma's conversational quality and verbosity. In its place, RLVR and RLHF were used via NeMo-RL and NeMo-Gym, with every sample trained both in reasoning mode and without. Length penalties in the GenRM environment limited the growth of responses. The dataset and training recipe are public. For deployment, the BF16 checkpoint was quantized to Q4_K_M with importance-matrix calibration on about one million RL rollout tokens; a DFlash drafter for speculative decoding is also included, which on an RTX 5070 reached 272.6 tokens per second versus 225.9 for Gemma 4 E4B with an MTP drafter, a 1.21x advantage claimed by Nvidia.
The model is distributed on Hugging Face, Ollama, LM Studio, Unsloth, and Jetson AI Lab, as well as as an Nvidia NIM and through the In-Game Inferencing SDK. In the cited demo, a Reachy Mini robot on an 8GB Jetson Orin Nano uses Cosmos 3 Edge for vision and Bolt for voice responses and tool calls. Nvidia calls it the third compact model in the family for ACE game characters.
The benchmark landscape
Nvidia grouped 21 benchmarks into five capability areas and states that Bolt leads 15 of the 21 rows in reasoning mode, with BF16 scores measured internally via NeMo-Gym. The best result cited is IHEval at 84.8 versus 67.0 for Gemma 4 E4B and 60.6 for Qwen3.5-4B. The company also reports average scores about 40% higher than Nemotron 3 Nano 4B and up to 80% fewer tokens generated in comparisons with active reasoning. Separately, Scale Labs reportedly ranked Bolt first among small models on MultiChallenge with 56.03%, according to an X post cited by the source.
The picture is not uniform: Qwen3.5-4B beats Bolt on Tau2-Bench (75.1 versus 55.1), RULER 64K (95.1 versus 87.0), LongBench v2 (59.9 versus 50.5), and BFCL-v3 (70.9 versus 68.5), while Gemma 4 E4B retains a narrow margin on EQ-Bench v3 (41.1 versus 39.4). Both the scores and the speed measurements remain to be independently verified; the source attributes these to Nvidia tests at batch size 1 with 3K input tokens and 50 output tokens.
