The decision maker
for voice commands.
Tars is a small local model. It takes one spoken command, splits it into steps, routes each step to the right app or tool, and picks the skill.
Illustration. The families and skills come from Tars's example catalog; the confidences are example values.
Three questions,
no text generation.
Tars never writes text. It makes a calibrated choice among options you supply, so every answer is one of your skills and comes with a probability. Your assistant acts when Tars is confident and falls back to its own rules when it is not.
Split
At each candidate word: does a new request start here? This is how one breath becomes several steps, even with no "and" between them.
Route
Once per step: which part of the assistant handles it, such as web, notes, messages or timers, or none if it isn't a request.
Pick the skill
Once per step: which skill within that family. The families and skills are data in a catalog file, not code.
Tars against the model
it was trained from.
Von 1.2, the base model, can answer questions over options but was never trained to split or route commands. Tars is the same model after fine-tuning.
Sentences fully right, by task
75 of the 480 sentences have a near-copy in Tars's training data. Leaving them out, Tars gets 61.2% overall.
Speed
Median time to plan a whole command, every split, route and skill question.
A single-step command takes about 130 to 230 ms.
Model size
The same size as Von: a full fine-tune of a ModernBERT encoder, not a bigger model. It runs offline on a laptop.
Benchmark: 480 spoken commands that each ask for 2 to 4 things in one breath, with three speaking styles kept out of training; latency measured in-process on an Apple M5 Pro.
Train it on your tools,
on your Mac.
A Mac app takes you from a list of your tools to a model that knows them. You bring the AI that writes the training labels; the training itself runs on your Mac.
Define your tools
List the families and skills your assistant has, with a few example phrasings each.
Generate labels
Your own AI provider, or any MCP client, writes messy and multi-step example commands with their labels.
Review
Go through the generated examples, fix or drop the wrong ones, and keep a test set the model never trains on.
Train and compare
Model Lab shows training as it runs and scores the new model against the old one on the same test set.
Export
Take the checkpoint to your assistant and serve plans from it over a small local HTTP server.
Runs locally
Planning happens on your Mac, with no network call. The only thing that leaves your Mac is label generation, and only to the AI provider or MCP client you choose.
Built on Von
Tars is a fine-tune of Von by Victor Hugo, an open-source decision model on a ModernBERT encoder. Von is licensed Apache-2.0, and so is Tars.
Open source.
Fine-tune your own.
The planner, the training pipeline and the Mac app are published under Apache 2.0. The Mac app needs Apple silicon and downloads about 3.4 GB the first time you train. Based on Von by Victor Hugo.