Models · Vinci by SimpleDirect®
Explore published models. Compare their files, licences, and evidence.
Check the exact release: files, licence, requirements, and evidence.
The download size is only the weights. Running the model needs extra memory.
For questions, writing, and conversation. Explore the Vinci family, from our first model to Piccolo and Bozza.
Open-weight release
A small model for everyday questions, writing, and conversation.
Open-weight release
A larger everyday model. See its test results and limitations before choosing it.
Vinci MLE is trained to help find out why an AI experiment went wrong, suggest a fix, or explain why nothing should change.
Compare Vinci MLE machine-learning engineering modelsEvaluation coverage differs by model. Valid held-out ML-engineering and external MLE-bench evaluations were not completed for 1.0. We plan to require those evaluations before releasing MLE 1.1.
A smaller model trained to help troubleshoot AI experiments. Needs less memory than 30B.
A larger model for the same work. Higher coding and tool-use test scores than 8B, but needs more memory.
Our largest MLE model, built on Devstral 2, with full BF16 weights and three GGUF formats under the Mistral Modified MIT licence.
Each model card compares measured general capabilities with its own parent model and documents gains, tradeoffs, and missing evaluations. These results do not establish performance on new ML-engineering tasks.
Why we’re building Vinci MLETrained to review security settings and suggest fixes for reported problems. Choose a size, then check its test results and licence.
Model cards and weights on Hugging Face
Compare Vinci Cyber defensive cybersecurity modelsFor human-reviewed defensive work.
A smaller model trained to suggest fixes for security settings—or leave them alone when no change is needed.
Broad CVE-repair improvement has not been demonstrated. CyberSOCEval results were below its parent, with regressions also observed on HellaSwag and GSM8K. The development panel informed checkpoint selection, and internal repair scoring is not independently qualified. It can miss issues or propose incorrect edits; human review remains necessary.
A larger model trained for the same security work. Needs more memory than 8B.
This model can change a clean configuration and introduce a security finding, as observed in its internal controls. PatchEval-Verified has not been measured on this checkpoint, and CyberSOCEval is pending. Internal repair scoring is not independently qualified; broad security superiority and safe autonomous operation are not established.
Our largest Cyber model, trained to review security settings and suggest fixes.
Cyber 123B can miss misconfigurations, change correct files or propose invalid fixes. It does not replace a scanner and does not know your environment’s compensating controls. The reported, bounded results do not establish general cybersecurity reliability, long-context performance or autonomous operation. Evaluate the actual format and serving setup you intend to use; results from Cyber 8B or 30B do not transfer to this model.
Flash is part of the Vinci record.
Historical model
Our first model, focused on legal writing and research in English and Canadian French.
Our next models are in development. We’ll share more when they’re ready.
In development
The next chapter in the Vinci model family.
In development
A new model in development for more demanding tasks.
In development
Working toward better, more versatile assistants for everyday tasks.
Experimental research
Prova is our experimental Mistral-based character-training study. Its public artifacts and limitations remain part of the research record; it is not recommended for production.
Explore Prova and the research artifacts