Foundation Model Program

Syntur’s Foundation Model Program is a long-term research initiative focused on developing original AI foundation models that Syntur designs, trains, evaluates, and eventually deploys. Our objective is to establish increasing technological independence by building our own model weights, training and inference infrastructure, evaluation systems, and, over time, proprietary datasets.

We intend to work toward training models from scratch, beginning with bounded research, reproducible evaluations, data preparation, and smaller-scale experiments. We are also evaluating the adaptation and fine-tuning of existing openly available models as preparatory work. These activities are distinct from developing an original foundation model.

Our long-term direction includes general-purpose intelligence and specialized models, including native coding models. Progress will be measured through documented technical milestones and evaluations rather than announced performance claims or speculative release dates.

Where the program stands

Not started means no work toward that item exists yet. Planned means intended and not built. Built offline means research tooling is implemented. It is not a completed run, and it has produced no approved result.

• Not startedTraining an original model from scratch. No original Syntur model weights exist, and no model has been trained from scratch. This is the program’s long-term goal. The program itself is underway as research, and the lines below are its current work.
• PlannedDatasets. No training data has been approved. Proprietary datasets are an over-time objective. Data from Varra conversations is not used for training.
• PlannedTraining and inference infrastructure for original models.
• Built offlineEvaluation. Offline tooling is implemented to compare an untuned model, a tuned one and the model Varra uses today on the same fixed prompts. No evaluation results have been reported, and none has been used to approve a model.
• Built offlinePreparatory fine-tuning. Offline tooling is implemented to prepare a dataset and train an adapter on a small open model, to test whether it can hold Varra’s response format. No fine-tuning run has been reported as complete, and no tuned model is approved or in use.

Varra runs on Anthropic’s Claude models. No Syntur model is in Varra, and none is offered to anyone. The wider research roadmap is on Research.

Original weights, adapters and licensed models

Original model weights

The parameters of a foundation model Syntur trains from scratch. Only these would make an original Syntur foundation model. None exist.

Licensed base models

Openly available models released by others under their own licenses. The program is evaluating two open model families as bases for preparatory fine-tuning research: OLMo 3 7B Instruct, from the Allen Institute for AI, under the Apache License 2.0, and Llama 3.1 8B Instruct, from Meta, under the Llama 3.1 Community License Agreement. Neither is a Syntur model.

Fine-tuned derivatives

A base model whose behavior has been adjusted by further training. A derivative combines Syntur’s contribution with a model trained by someone else, so it is not an original Syntur foundation model.

Adapter weights and other modifications

An adapter is a small set of additional parameters trained on top of a base model to adjust its behavior. Syntur may own the adapters and other modifications it creates. Their use, distribution and commercialization, alone or with the base model, remain subject to the base model’s license and any other applicable restrictions. Apache 2.0 lets a licensee set its own terms for its modifications, provided its use of the original work still complies with that license. The Llama 3.1 license treats derivatives a licensee makes as the licensee’s, subject to Meta’s ownership of the Llama materials, and attaches conditions to redistribution, including attribution, model naming and an acceptable use policy.

Licensed training data

Data used under its own terms. Training data must be written by people or used under a license that permits it, and must be reviewed before use. Outputs generated by AI providers’ models are not used as training data without written confirmation that the use is permitted.

How progress will be reported

• By milestone and evaluation. Each step is reported when it is documented and measured: what was built, what was evaluated, on what, and with what result. There are no announced performance claims and no release dates.
• Measured before it is trusted. Every compared run uses the same fixed prompts, the untuned model is measured before any tuning, and a result on a small prompt set is reported as a smoke test, not a measurement.
• Behavior, not facts. Facts that matter come from retrieval or verified reference data, not from what a model remembers.
• Tested before anyone relies on it. Before any potential user-facing deployment, a model would go through documented evaluation, controlled testing and safeguards.

The model catalog stays unpublished until Syntur releases a model of its own.