Fort makes the call and tells you how sure it is
Fort-1 picks one of your options and puts a probability on the call, so your software knows when to act and when to ask a person. Two small open models that run on your own hardware.
A quiet film that stays with you for days after the credits roll.
- Apositive99%
- Bnegative1%
DecisionA · positive99%
28 ms
Fort is our first model family, built to make decisions, not conversation.
What is a System One model?Only your options
Fort reads your options and returns one of them, with its probability. It writes no free text, so it cannot ramble or invent an answer.
Measured confidence
Fort is built to be right about as often as it says it will be. On SST2, Fort-1 2B's calibration error is 0.012, against 0.026 for Jev.
Small, open and local
Open weights under Apache-2.0, for MLX and PyTorch. At 4-bit with MLX, Fort-1 0.8B is a 0.44 GB download and decides in 0.7 GB of memory. Nothing leaves the machine.