NVIDIA released Nemotron 3.5 Bolt E4B 4GB on October 23, 2026, for conversational agents with limited memory. According to NVIDIA, the model supports up to 64K context, multi-turn tool use and configurable reasoning. Sources: NVIDIA
NVIDIA post-trained Google’s Gemma-4-E4B-it to build the model, which has 8 billion total parameters and 4 billion effective parameters. The release takes text input and produces text output, uses Q4_K_M quantization and has a companion BF16 model. Sources: NVIDIA
NVIDIA reported generation speeds of 133.4 tokens per second on an RTX 5070 and 15.5 tokens per second on a Jetson Orin Nano. Both tests used llama.cpp with CUDA, batch size one, no cached context, approximately 3,000 input tokens and 50 output tokens. These figures are vendor measurements, not independent benchmarks. Sources: NVIDIA
The 4GB name describes a target workload footprint rather than a universal memory ceiling. NVIDIA measured peak memory of 3.5GB on the RTX 5070 and 4.2GB on the Jetson Orin Nano. Memory reached 5.0GB across its listed configurations. Sources: NVIDIA