Vinci MLE 1.0 introduces 8B and 30B open-weight research models developed by SimpleDirect® in Canada, building on IBM Granite 4.1 and released under Apache-2.0.
Their specialization focuses on inspecting experiment evidence, diagnosing problems, and making bounded changes.
Training also includes preserving a correct setup and stopping when more evidence is needed.
Both models include full merged weights, their tokenizers, chat templates, loading examples, and detailed model cards.
GGUF distributions are also provided, with their own conversion records and runtime checks; BF16 benchmark results do not establish GGUF capability retention.
What this release establishes
The cards report training scope and general-capability comparisons. A valid held-out MLE-task evaluation was not completed for 1.0.
For 1.1, we intend to make held-out MLE evaluation and an external MLE benchmark suite release-gating measurements, alongside general capability-retention evaluations.
The weights alone are not the full Vinci MLE agent harness. Applications must supply and validate tools, workspace permissions, and an isolated execution environment.
Explore the models or read George’s post on why we’re building Vinci MLE.
George Pu is the founder and CEO of SimpleDirect, an independent Canadian AI lab developing research, models, technologies, and products under Vinci.
