Fort-1 2B
A small open decision model on Qwen3.5-2B. Give it a text and your options: it returns one option with a probability, the probability of yes, or a place on an ordered scale, read from the model's probabilities with no generated text. Apache-2.0; runs on MLX and PyTorch.
A rank-16 LoRA whose targets blend a Qwen3.6-35B-A3B teacher's probabilities with labels from Teximal's annotators. Choice handles up to 26 options; longer label lists are read by name. teximal serve answers Decisions-API-shaped requests. Sibling Fort-1 0.8B (built for speed) and 4-bit MLX builds of both. The numbers are Teximal's own runs of Dhruv Mehra's jevbench (github.com/dhruvmehra/jevbench), which is a different suite from JevBench; Banking77 is adapted, not zero-shot. Long label lists are overconfident until a temperature is fitted. Text only, mostly English, not for maths. The maker says the training data includes research-only and share-alike sets.
What it decides
- choice — picks one option from a set
- noul — answers a yes/no question with one probability
- score — places the input on an ordered scale
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 1.88B |
| Base model | qwen/qwen3.5-2b |
| Maker | Teximal |
| Released | 2026-10-03 |
| License | apache-2.0 |
| Reported accuracy | 93.8% |
| Reported latency | 110 ms median per decision on an Apple M1 Max (MLX, one at a time) |
Get the weights
pip install systemonemodels
systemone pull teximal/fort-1
The files are served from the maker's Hugging Face repository, teximal/fort-1-2b, and verified against the checksums recorded here.