Apriel-v1.5-15B-Thinker is an open-weights multimodal reasoning model developed by ServiceNow. It is designed to deliver performance comparable to significantly larger models—such as DeepSeek-R1-0528 and Gemini 2.5 Flash—while remaining deployable on a single GPU. Released under the MIT license, the model emphasizes computational efficiency and accessibility for enterprise and research applications.
The model's architecture is built upon the Pixtral-12B foundation, utilizing a technique known as depth upscaling to expand the network from 40 to 48 decoder layers. This process increases the total parameter count to approximately 15 billion, providing the capacity needed for complex reasoning tasks without requiring a full training run from scratch. It features a multimodal design with a vision encoder and projection network, enabling it to process both text and image inputs.
Apriel-v1.5-15B-Thinker is notable for its midtraining pipeline, which relies on staged continual pre-training (CPT) and high-quality supervised fine-tuning (SFT) rather than reinforcement learning or preference optimization. The training curriculum includes billions of tokens focused on mathematical reasoning, coding challenges, scientific discourse, and multimodal tasks such as document and chart understanding. This data-centric approach results in strong performance on benchmarks like AIME 2025 and GPQA Diamond, as well as specialized enterprise benchmarks like Tau2 Bench Telecom.