ServiceNow AI Unveils Apriel-1.5-15B-Thinker

In a significant advancement for AI research, ServiceNow AI has introduced Apriel-1.5-15B-Thinker, a 15-billion-parameter multimodal reasoning model that achieves frontier-level performance while remaining deployable on a single GPU. This release challenges the prevailing assumption that high-performance AI requires extremely large models and vast computing resources. By combining innovative training strategies with an emphasis on efficiency, ServiceNow AI demonstrates that cutting-edge AI can now be more accessible to organizations with limited hardware infrastructure.

Apriel-1.5-15B-Thinker excels across multiple domains, including text, code, and vision, offering a highly versatile model capable of handling complex reasoning tasks. Its design is particularly relevant for enterprises or researchers who need robust AI capabilities without investing in clusters of expensive GPUs. By providing open weights, training recipes, and evaluation protocols under the MIT license, ServiceNow AI also promotes transparency and reproducibility, critical principles in modern AI development.

---

Key Highlights

Frontier-Level Performance: Despite being approximately 8–10 times smaller than some state-of-the-art models like DeepSeek-R1-0528, Apriel-1.5-15B-Thinker achieves an Artificial Analysis Intelligence (AAI) Index score of 52, demonstrating competitive reasoning and problem-solving ability across a wide array of tasks. This metric positions the model at the frontier of multimodal AI performance.

Single-GPU Deployability: One of the most striking features of Apriel-1.5-15B-Thinker is its ability to run on a single GPU, opening up possibilities for researchers in resource-constrained environments. Traditionally, frontier-level AI models have required multiple high-end GPUs or large-scale cloud infrastructures, making them inaccessible to many academic institutions, small companies, and independent developers.

Open Weights and Transparent Development: ServiceNow AI’s decision to release the model weights and training methodology fosters a collaborative research environment. Other AI researchers can reproduce results, benchmark the model in new scenarios, or fine-tune it for specialized applications, further accelerating AI innovation.

Comprehensive Multimodal Capabilities: Apriel-1.5-15B-Thinker performs strongly on benchmarks such as AIME 2025 (≈88), GPQA Diamond (≈71), LiveCodeBench (≈73), Instruction-Following Benchmark (62), and Tau-squared Bench Telecom (68). These scores indicate its ability to handle reasoning across text, numerical, and visual domains, making it a true multimodal model.

ServiceNow AI employed a three-stage training approach to achieve this remarkable combination of performance and efficiency:

Depth Upscaling: The model’s architecture was enhanced by increasing depth, expanding reasoning capabilities without needing to train from scratch, which saved considerable computation.

Staged Continual Pretraining: This stage focused on foundational understanding of text and vision, followed by advanced visual reasoning. Targeted synthetic datasets were used to improve spatial awareness, compositional understanding, and fine-grained perception.

1. Supervised Fine-Tuning: Finally, the model was fine-tuned with high-quality instruction-response pairs. These included reasoning traces covering mathematics, coding, science, and tool usage, allowing the model to perform structured reasoning tasks efficiently.

Interestingly, Apriel-1.5-15B-Thinker was trained without reinforcement learning or preference optimization. This approach isolated the benefits of data-centric continual pretraining, highlighting the effectiveness of thoughtful data preparation and staged learning strategies.

---

Benchmark Performance

Across multiple industry-standard and academic benchmarks, Apriel-1.5-15B-Thinker demonstrates impressive results:

• AIME 2025: ≈88

• GPQA Diamond: ≈71

• LiveCodeBench: ≈73

• Instruction-Following Benchmark: 62

• Tau-squared Bench Telecom: 68

These results place it on par with larger models like Gemini-2.5-Flash and Claude Sonnet-3.7, despite its much smaller size and single-GPU requirement. This makes it an appealing option for scenarios where computational efficiency is essential without compromising on accuracy or reasoning depth.

---

The versatility of Apriel-1.5-15B-Thinker opens doors to a wide range of applications:

• Enterprise Automation: Tasks such as document analysis, automated coding, and knowledge extraction can be handled efficiently with this model.

• Research and Academia: Universities and small research labs can now experiment with frontier-level AI without needing access to supercomputing resources.

• Software Development Assistance: The model’s multimodal understanding allows for better code completion, debugging support, and intelligent documentation generation.

• Vision-Language Reasoning: From image captioning to advanced visual question-answering, the model supports complex multimodal reasoning tasks.

---

Researchers and developers can access Apriel-1.5-15B-Thinker on Hugging Face (ServiceNow-AI/Apriel-1.5-15b-Thinke). The open-access approach allows the community to reproduce results, fine-tune for specific domains, and further explore the potential of efficient multimodal reasoning.

---

Apriel-1.5-15B-Thinker exemplifies how innovative training methodologies can produce high-performance AI models without massive computational costs. Its open weights and transparent development process strengthen the open-source AI ecosystem, making frontier-level multimodal reasoning accessible to organizations and researchers with limited resources. By balancing efficiency with capability, ServiceNow AI sets a new standard for responsible, resource-conscious AI development.