Arcee.ai Releases Virtuoso-Small: A Compact 14B Parameter Model for Business-Oriented Generative AI

Arcee.ai has unveiled the first public release of its Virtuoso series, introducing Virtuoso-Small, a 14 billion parameter model designed to bring high-performance generative AI capabilities to organizations and developers in a compact form. This model is an excellent entry point for tasks requiring instruction-following, complex reasoning, and business-oriented applications. Larger variants, Virtuoso-Medium and Virtuoso-Large, are also available via API at models.arcee.ai.

Key Technical Changes and Why They Matter

Virtuoso-Small stands out for its balance of performance and efficiency. Here’s a breakdown of what makes it noteworthy:

• Compact and Efficient: Despite having 14 billion parameters, Virtuoso-Small is optimized for smaller hardware configurations. This means developers can deploy it on less powerful machines without compromising quality.
• Business-Oriented: The model is tailored for enterprise use cases such as customer support, content creation, and technical assistance. It’s designed to meet the specific demands of modern businesses.
• Scalable Ecosystem: Virtuoso-Small is part of a family that includes Virtuoso-Medium and Virtuoso-Large. This allows organizations to start with a smaller model and scale up as their needs evolve.

Technical Details

• Model Size and Parameters:

  • 14 billion parameters (15B params)
  • Tensor Type: BF16
  • License: Apache-2.0

• Base Model:

  • Finetuned from Qwen/Qwen2.5-14B

• Deployment Options:

  • Local Deployment: The model can be deployed on local hardware, making it suitable for environments where data privacy and control are critical.
  • API Access: For larger-scale or more demanding applications, the model is available via API at models.arcee.ai.

Performance Benchmarks

Virtuoso-Small has been evaluated on several benchmarks to gauge its performance. Here’s a summary of the key metrics:

• Average Score: 39.43
• IFEval (0-Shot): 79.35
• BBH (3-Shot): 50.40
• MATH Level 5 (4-Shot): 34.29
• GPQA (0-shot): 11.52
• MuSR (0-shot): 14.44
• MMLU-PRO (5-shot): 46.57

These scores indicate that Virtuoso-Small performs well in a variety of tasks, from zero-shot learning to complex reasoning and multi-shot evaluations.

Model Ecosystem

Virtuoso-Small is part of a broader ecosystem with several related models:

• Adapters: 2 models
• Finetunes: 4 models

These additional models can be used to further customize Virtuoso-Small for specific use cases or to enhance its performance in particular domains.

Community and Support

While Virtuoso-Small is not currently deployed by any Inference Provider, the community can request support through [Hugging Face's InferenceSupport space](https://huggingface.co/spaces/huggingface/InferenceSupport/discussions/new?title=arcee-ai/Virtuoso-Small&description=React%20to%20this%20comment%20with%20an%20emoji%20to%20vote%20for%20%5Barcee-ai%2FVirtuoso-Small%5D(%2Farcee-ai%2FVirtuoso