Fine-Tuned Small AI Models for Agentic Workflows

No. Every model is an INT8 ONNX graph that runs on the CPU execution provider of onnxruntime. A small GPU or an edge device works too, but nothing requires one. Each model page lists its size and parameter count.

No. You download the model and run it yourself, so the text and documents it reads stay on your infrastructure. Running it makes no network call of any kind.

A model directory: the ONNX graph, its tokenizer, a serving.json with the labels and the decision rule, and a model card with sample inputs. Each file comes with its size and SHA-256 so you can verify the copy.

Each model page shows the base model, what it was trained on, the use cases it was built for and real example results — what this build answered on held-out rows it never saw in training. Read those before integrating, then test it on your own data after you download it.

Indonesian and English, each as its own build of the model. The vocabulary is trimmed to those two languages, which is most of why the files are small. A third language is a separate model with its own training data; ask us.

Book a 30-minute call at https://calendly.com/hello-intura/30min and you get an engineer on it, or write to [email protected]. Neither is needed to try a model: during early bird every model is $9 to download.