Fort-1 2B
Fort picks one of your options for a text and tells you how sure it is. It never writes free text: every answer is one of your options, with a probability you can act on: take the sure answers, send the rest to a person or a larger model.
Fort-1 2B is the larger model, for accuracy. Fort-1 0.8B is the small model, for speed.
| Parameters | 1.88B, fine-tuned from Qwen3.5-2B |
| Download | 3.8 GB (bf16); 4-bit MLX: 1.08 GB |
| Memory while deciding (MLX) | 4.2 GB (bf16); 1.6 GB (4-bit) |
| Context window | 262,144 tokens |
| Speed | 79 to 214 ms per decision (jevbench medians, Apple M1 Max) |
| License | Apache-2.0 |
Use
pip install teximal
teximal run fort-1-2b "I was charged twice this month." --options billing,technical,sales
teximal serve fort-1-2b # Jev's request format (OpenRouter's Decisions API): switching is a change of URL
from teximal import Fort
fort = Fort("teximal/fort-1-2b")
route = fort.task({"billing": "invoices, payments, refunds", "technical": "bugs, outages, error messages",
"sales": "plans, upgrades, pricing"}, question="Which team should handle this?")
d = route.decide("I was charged twice this month.")
print(d.choice, d.confidence) # the option and its probability
many = route.decide_many(["Refund please.", "The app crashes on start."])
A question is a choice (up to 26 options, each with a probability; longer label lists are read by name), yes or no (the probability of yes), or a score on an ordered scale. Docs: teximal-cli.
Results
jevbench, 500 examples per task: accuracy in %, calibration error (ECE, lower is better) and median milliseconds per decision.
| Model | SST2 | AG News | Banking77 | ECE SST2 / AG / B77 | ms SST2 / AG / B77 |
|---|---|---|---|---|---|
| Fort-1 2B | 93.8 | 86.8 | 76.6* | 0.012 / 0.092 / 0.182 | 110 / 214 / 79 |
| Fort-1 0.8B | 93.2 | 87.2 | 74.8* | 0.013 / 0.063 / 0.178 | 28 / 44 / 51 |
| Jev (jev-1.13, TypeSafe API) | 95.4 | 84.3 (85.8†) | 76.4 | 0.026 / 0.112 / 0.125 | 376 / 381 / 389 |
| Claude Sonnet 5 (Anthropic API) | 95.6 | 89.6 | 77.4 | not reported | 2,076 / 2,014 / 1,995 |
| GPT-5-mini (OpenAI API) | 95.0 | 80.2 | 73.6 | not reported | 1,385 / 1,312 / 1,312 |
- Fort's times: one decision at a time on an Apple M1 Max (MLX). The API models' times include the network.
- Fort never trained on SST2 or AG News, only on similar tasks: IMDB review sentences and HuffPost headlines labeled by Teximal Lab.
- *Banking77 is adapted, not zero-shot: Fort trained on bank messages Teximal Lab wrote from the 77 intent names, never on real Banking77 messages (one scored message also appears inside a training message). Scored with jevbench's instruction, as every model was; with a banking instruction, Fort-1 2B scores 77.8 and Fort-1 0.8B scores 74.6.
- †85.8 with jevbench's newer AG News label definitions.
Calibration: one temperature fitted on a few hundred labeled examples of your task
(teximal eval does it) takes ECE to 0.018 / 0.060 / 0.035 (Fort-1 2B) and 0.020 / 0.062 / 0.038 (Fort-1 0.8B), on 300 held-out examples per task.
More results
| Test | Fort-1 2B |
|---|---|
| Decision exam: 600 tasks written by Teximal Lab, never trained on | 80.3% (ECE 0.035) |
| People's 1-to-5 ratings, held out (urgency, frustration, sentiment strength, toxicity, stars) | 52 to 76% exact, 92 to 98% within one level |
| General knowledge, 4 options | 92.5% |
| Reading comprehension, 4 / 16 options | 96.8% / 93.5% |
| Assistant requests in 51 languages (MASSIVE) | 78.1% |
| Options shuffled or renamed, typos, lower case, a distracting sentence | at most 1.7 points lower |
| A text that claims the answer or orders a letter | at most 1% of decisions moved |
| Long context (window: 262,144 tokens) | 90 to 95% with the deciding text inside about 32,000 tokens; 3 of 3 at about 129,000 |
Limitations
- Not a math model: it answers in one step, without working.
- Mostly English; tweet sentiment in four African languages is weak (AfriSenti 49.9%).
- Text only.
- Long label lists read by name (such as Banking77) are overconfident until a temperature is fitted.
- Instructions planted in a text rarely move its answer, but treat untrusted text as data.
Training
Fine-tuned from Qwen/Qwen3.5-2B with LoRA (rank 16). The targets blend the probabilities of a larger teacher, Qwen3.6-35B-A3B (Apache-2.0), with people's labels from Teximal Lab, collected with hidden checks.
| Data | Used for | License |
|---|---|---|
| MASSIVE | assistant requests in 51 languages | CC BY 4.0 |
| SQuAD 2.0 | reading, answerability and knowledge questions | CC BY-SA 4.0 |
| CLINC150 | intents and intent names | CC BY 3.0 |
| DBpedia14 | entity types | CC BY-SA 3.0 |
| SNLI, MultiNLI | entailment | CC BY-SA 4.0; MultiNLI's own mixed terms |
| Amazon reviews (amazon_polarity) | sentiment; 1-to-5 star ratings by Teximal Lab | as published on Hugging Face |
| Civil Comments | toxicity; 1-to-5 toxicity ratings by Teximal Lab | CC0 1.0 |
| HuffPost News Category Dataset | news headlines, topics labeled by Teximal Lab | CC BY 4.0 |
| IMDB Large Movie Review Dataset | review sentences, sentiment labeled by Teximal Lab | research dataset |
| Teximal Lab | critic-style sentences and bank customer messages, written from scratch | Teximal |
Intended use
Routing, triage, moderation queues, tagging and form checks: decisions over your own options, where a calibrated confidence decides what to automate and what to send to a person. Not for decisions with legal or similarly significant effects on people without human review.
License
Apache-2.0; the Qwen3.5 base model's notice is in NOTICE.
Citation
@misc{teximal2026fort1,
title = {Fort-1: calibrated decision models},
author = {Teximal},
year = {2026},
url = {https://huggingface.co/teximal/fort-1-2b}
}