NVIDIA ships Nemotron 3.5 Bolt E4B for CPU agents

NVIDIA released Nemotron 3.5 Bolt E4B on October 23, a compact agent-tuned model built on Gemma 4 E4B with 8B total and about 4B effective parameters and a 64K context. Running the 4-bit GGUF on an 8-vCPU Xeon with no GPU yielded about 11 tokens/s generation and 95% correct tool arguments.

• Model is post-trained from Gemma 4 E4B with RL for tool calling
• Q4_K_M GGUF is 6.18 GB on disk with a 64K context
• On 8 Xeon vCPUs: 194.6 tokens/s prefill, 10.9 tokens/s generation
• Argument accuracy was 57 of 60 with thinking mode off