ServiceNow releases Apriel-Nemotron-15B-Thinker to match 32B models at half memory

ServiceNow has introduced Apriel-Nemotron-15B-Thinker, a 15-billion parameter model in the Apriel SLM series that matches or exceeds the performance of larger 30-32 billion parameter baselines like o1-mini, QWQ32B, and EXAONE-Deep-32B while using approximately half their memory footprint.

The model is built on a four-stage training pipeline: base model upscaling from a 12B backbone, continual pre-training on reasoning traces, supervised fine-tuning on high-quality reasoning data, and reinforcement learning using Group Relative Policy Optimization (GRPO).

Evaluations across seven enterprise-focused benchmarks and five academic reasoning tasks demonstrate that the model fits within the memory capacity of a single H100 or dual consumer GPUs while delivering state-of-the-art results for its size.

This approach addresses the gap for models small enough to run on limited hardware yet smart enough for complex multi-step reasoning, tool invocation, and domain-specific enterprise tasks.