Skip to Content
Models and evidence

Models and evidence

Use the canonical Vinci model catalogue  to check current release availability, then the individual model card for loading instructions, lineage, licence and evaluation scope. This guide is an entry point, not a replacement for those records.

The release snapshot below follows SimpleDirect’s public model records reviewed on September 28, 2026. A downloadable model is not automatically served through the hosted API, supported by Vinci Code CLI, or suitable for production use.

Before you download

  1. Choose a specific size and read its model card, including limitations.
  2. Check its licence and the licence of any derivative or quantization you use.
  3. Follow that card’s runtime and hardware instructions. Model weights alone do not supply an application, execution harness, tools or permissions.
  4. Evaluate representative tasks in your own environment. Published results apply only to the tested model, format, population and setup.

Vinci MLE 1.0

Research models for ML-engineering workflows, released September 28, 2026.

A valid held-out MLE-task evaluation was not completed for 1.0. Published weights and other benchmark results do not establish proven autonomous ML engineering.

ArtifactLicenceModel card and weightsRelease information
Vinci MLE 8B 1.0Apache-2.08B model card 8B record 
Vinci MLE 30B 1.0Apache-2.030B model card 30B record 

MLE 123B is not listed as released or downloadable in this September 28 snapshot. Check the catalogue  for subsequent changes. The MLE release record  explains the release scope and evaluation limitation.

Vinci Cyber 1.0

Specialist models for defensive cybersecurity, released September 22, 2026. These are fine-tuned models, not models pretrained from scratch.

ArtifactLicenceModel card and evaluation evidenceRelease information
Vinci Cyber 8B 1.0Apache-2.08B model card 8B record 
Vinci Cyber 30B 1.0Apache-2.030B model card 30B record 
Vinci Cyber 123B 1.0Modified MIT (Mistral AI); see the release licence 123B model card 123B record 

Evidence is size-specific. Read each card; do not apply one size’s findings to the whole family. The pinned 123B aggregate evaluation  reports BF16-parent-plus-adapter measurements on an internal panel. It does not establish a task rerun on merged weights or GGUF, independent security certification, or safe autonomous deployment. The Cyber release record  summarizes the mixed findings and their limitations.

Research and earlier artifacts

The research record  includes numbered technical reports, corrections and negative findings. Earlier Piccolo, Bozza and Prova artifacts remain in the catalogue . Prova is experimental research; it is not a production recommendation.

For software that acts on a repository, start with Vinci Code CLI. For an existing hosted-service integration, use the API reference; its service routing names are not downloadable model identifiers.

Last updated on