Applied research
Lapis Labs
Smaller models. Better-defined decisions.
We explore more efficient assistants: models with a specific task, well-chosen tools and more processing on your own device.
Featured project ·
Lapis SAL
A teacher, a compact distilled student and an MLX release for English and Spanish tool selection, with an explicit option to abstain.
Read the publication- Open project
- Apache 2.0
- PyTorch / MLX
- English and Spanish
“Add a meeting tomorrow”
Publications
Technical announcements, methods and results. Research, separate from updates to the app.
Open models
Checkpoints, runtimes and instructions published on Hugging Face. Apache 2.0.
lapis-sal-150M
Student · research and fine-tuning
PyTorch FP32
lapis-sal-421M
Teacher · inference and distillation
PyTorch FP32
lapis-sal-150M-MLX
Student · Apple Silicon
MLX FP16 · load-time INT8
A specific task. An explicit evaluation.
Tool selection
We study the decision before an action: choosing a tool, or recognizing that none is appropriate.
Distillation and efficiency
We explore transferring a teacher’s decisions into a more compact encoder.
Local inference
We prepare runtimes for focused models on Apple Silicon, using MLX and reduced-precision formats.
Publishing means explaining the limits, too.
Each publication documents the architecture, available formats and the context behind its results. Checkpoints and instructions make the work testable; evaluation needs to be repeated with data representative of each application.
SAL results come from a private, synthetic benchmark. They do not guarantee general performance or validate actions executed by an assistant.
These publications describe research. They do not mean that the models or features are available in the app.