Developed by ServiceNow’s Language Models Labs (SLAM Labs), Apriel-v1.6-15B-Thinker is a 15-billion parameter multimodal reasoning model designed for enterprise-grade efficiency. It serves as an iteration in the Apriel series, focusing on advancing frontier-level reasoning capabilities while minimizing the computational footprint. The model is specifically optimized to balance high intelligence scores with token efficiency, reportedly reducing reasoning token usage by over 30% compared to its predecessor.
The model's architecture is the result of depth upscaling, expanding a 12-billion parameter base to 15 billion parameters by increasing the layer count from 40 to 48. Its training methodology follows a three-stage progressive process: initial depth upscaling, followed by staged continual pretraining (CPT) on a curated mixture of high-quality reasoning and synthetic data, and a final post-training phase incorporating supervised fine-tuning (SFT) and reinforcement learning (RL).
Apriel-v1.6-15B-Thinker is capable of handling both text and image inputs, with specific strengths in mathematical reasoning, code generation, and complex document understanding. It was trained using NVIDIA DGX Cloud infrastructure and GB200 Grace Blackwell Superchips. By maintaining a 15B parameter scale, the model is designed to be deployable on a single GPU, making advanced reasoning more accessible for production workloads in regulated industries.