Today, we’re releasing Vinci MLE 123B 1.0, our largest model for machine learning engineering. Its full BF16 weights and GGUF versions are now public on Hugging Face.
We developed it at SimpleDirect® in Toronto, building on Mistral’s Devstral 2. It joins the 8B and 30B models in the Vinci MLE family.
The direction is practical: help with the work of building AI. Investigate why an experiment failed. Read the evidence. Suggest a limited fix, preserve a working setup, or explain what information is missing before making a change.
That is what we trained MLE to work toward. Developers can now download the model, connect it to their own tools, and evaluate it in their own environments.
The model cards document the measured general-capability results, including gains and tradeoffs. Coding evaluation was not completed for 123B, and valid held-out ML-engineering and external MLE-bench evaluations were not completed for 1.0. Those MLE evaluations are part of our planned 1.1 work.
The downloads include full BF16 weights and Q4_K_M, Q5_K_M, and Q8_0 GGUF formats. The GGUF card explains the runtime checks and their limits; these checks do not establish equivalent task performance across formats.
123B uses Mistral’s Modified MIT licence, with its preceding-month revenue condition. The 8B and 30B models use Apache-2.0. Please read the licence for the model you choose.
For me, this release is another step toward making Vinci useful to the people building AI: models they can inspect, run, and test against the work that matters to them.
Explore Vinci MLE 123B 1.0 or download the GGUF versions.
Cover: conceptual brand illustration, not a model run or evaluation result.
George Pu is the founder and CEO of SimpleDirect, an independent Canadian AI lab developing research, models, technologies, and products under Vinci.
