English–Spanish tool-routing models: a 421M teacher, a distilled 150M student, and an MLX release for Apple Silicon.
AI & ML interests
On-device AI, edge inference, small language models, tool routing, MLX, multimodal models, model quantization
Recent Activity
Organization Card
💎 Lapis Labs
Small models. Practical AI.
We build efficient models that help AI assistants decide which tool to use—or when no tool is needed.
Our Lapis SAL family focuses on English–Spanish tool routing, with a larger teacher, a distilled student, and an MLX release for local inference on Apple Silicon.
Why tool routing?
Assistants often need to choose between searching the web, checking a calendar, sending a message, or answering directly.
A dedicated routing model can handle this decision before the request reaches a larger language model or an external service. This supports local processing and gives developers explicit control over the available tools.
Available models
| Model | Purpose |
|---|---|
| Lapis SAL 150M | Distilled PyTorch student for research, further training, and export. |
| Lapis SAL 421M | Larger PyTorch teacher used in the distillation workflow. |
| Lapis SAL 150M MLX | Apple Silicon release with FP16 inference and optional INT8 quantization at load time. |
The student’s 150M name identifies its encoder family. The complete routing model contains approximately 164M learned parameters.
Supported tools
Calculator · Calendar · Clock · Code · Contacts · Health · Location · Mail · Maps · Messages · Reminders · Weather · Web Search
A NONE option covers requests that do not require an available tool. Developers supply the tool descriptions at inference time.
Training and evaluation
The student learns from a larger teacher through knowledge distillation. The workflow includes supervised adaptation, calibration, and deployment export.
On our private, synthetic, held-out benchmark of 1,000 English–Spanish examples:
- Teacher: 96.0% routing accuracy.
- MLX student: 95.1% with both FP16 and optional INT8 inference.
These results describe this benchmark and do not guarantee performance on other datasets or real-world traffic. See each model card for evaluation details and limitations.
Get started
Choose a model above and follow its README for installation and usage.
These releases use a custom routing runtime. They select tools; they do not execute them, generate conversational responses, or provide medical advice.
English & Spanish · Knowledge distillation · Local inference · Apache 2.0
datasets 0
None public yet