Open-weight AI you can run yourself.
Open-weight AI makes model parameters available for download under the terms of a particular licence. Vinci by SimpleDirect includes specialist language models with public model cards, weights and release-specific disclosures. Start with those artifacts before choosing a deployment.
Available weights are a starting point.
Downloadable weights let you run a model outside a hosted assistant. Rights to modify or redistribute it depend on the licence. Open weights do not automatically mean open-source training code or public training data.
- Read the licence for the exact model and its upstream foundation.
- Keep the model version, file format and quantization with your deployment record.
- Access to weights is not proof of accuracy, safety or suitability for a regulated workflow.
Self-hosting gives you choices and responsibilities.
Local, on-premise and private-cloud deployment can help an operator control infrastructure and access. The complete system still includes runtimes, prompts, logs, tools and network connections.
- Choose a runtime and file format that support the artifact you download; memory and context requirements vary.
- Review telemetry, logging, tool calls and external services before using sensitive data.
- Canadian ownership does not guarantee Canadian data residency. Verify the actual hosting and data paths of your configuration.
Match the model to a task you can check.
Specialist models are intended for particular workflows, not every technical problem. A benchmark result or completed run does not replace reviewing the output against the task requirements.
- Vinci Cyber targets defensive cybersecurity. Its published evaluation scope does not establish complete security or general code-audit capability.
- Vinci MLE targets ML-engineering workflows; MLE 1.0 lacks a valid held-out ML-engineering task evaluation.
- Plan a bounded trial, define acceptance criteria in advance and keep human review for consequential decisions.
Choose a family, then inspect the artifact
Family pages explain intended work. Model cards and evaluation records carry the release-specific facts; source materials may be in English.
Vinci Cyber defensive cybersecurity models
Compare the family, foundations and limits before a security trial.
Vinci MLE machine-learning engineering models
Explore intended workflows and the evaluation gaps to account for.
Choosing model files
Match a download to your runtime and hardware budget.
Reading model evaluations
Check the population, comparison and what a result leaves unproven.
Start with a small, reviewable trial.
Choose a concrete task, record the exact artifact and check the result before expanding its role.