Small Specialized Language Model Experts
Our models are small but mighty!
Ranging from 1-22B parameters, we have developed over 150 small, specialized CPU- and laptop friendly language models in Hugging Face.
Our latest innovation is the SLIM model portfolio (Structured Language Instruction Models) - the world's first function-calling structured output small language models developed to be used in multi-model, Agent-based workflows.
We custom-train and optimize models for specialized use cases in data-sensitive, highly-regulated industries such as financial services, legal and insurance to run in private cloud or on-prem.
Find us in Hugging Face! ๐ค
WHY DEVICE-OPTIMIZED MODELS
Match the model to the hardware you already use.
Apple Silicon
Use Metal-ready formats to get efficient local inference on M-series Macs.
AMD Ryzen AI
Pair ONNX and Vitis-ready models with AMD NPUs and integrated GPUs.
Intel Core Ultra
OpenVINO-aligned models help use Intel CPU, GPU, and NPU resources together.
NVIDIA GPUs
CUDA-ready runtimes unlock high-throughput inference for GPU workloads.
Qualcomm Snapdragon
QNN-ready options bring efficient on-device AI to Windows on ARM systems.
DRAGON Models
Production-grade RAG-optimized 6-7B parameter models - "Delivering RAG on ...". Fine-tuned for question-answering, fact-based uses in RAG workflows. Uses the context provided to answer yes/no and multiple-choice questions. Trained to reduce hallucinations.
SLIM Models
Function-calling, structured output models for classifying and clustering tasks. Designed to be used in an AI Agent workflow either alone or stacked together. 10+ Structured Language Instruction Models (SLIMs) for almost every classifying task.
Industry BERT Models
Industry and specialized domain finetuned BERT embedding models. Specialized domains include: Insurance, SEC documents, Contracts & Asset Management.
SLIM GGUF
Quantized GGUF "tool" implementations of SLIM Models. Provide "gguf" and "tool" versions of many SLIM, DRAGON, and BLING models, optimized for CPU deployment. GGUF Generative Model class - support for Stable-LM-3B, CUDA build options, and better control over sampling strategies.
BLING Models



