# Axel Montzamir (Okura66): kahn1

> Open 4B decision model by Axel Montzamir: reads a text state and answers typed choice, score and noul questions in one forward pass, taking probabilities from the option tokens with permutation debiasing. Apache-2.0 weights; the 3B sibling is research-only.

- Page: https://systemonemodels.ai/okura66/kahn1
- API: https://api.systemonemodels.ai/v1/models/okura66/kahn1
- Download: `pip install systemonemodels && systemone pull okura66/kahn1`

## Facts

| | |
|---|---|
| Maker | Axel Montzamir (Okura66) (https://systemonemodels.ai/okura66) |
| Decides | choice, score, noul |
| Architecture | kahn1 |
| Base model | qwen/qwen3.5-4b |
| Parameters | 4.0B |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-09-25 |
| Latest version | 2026.09 |

## Reported evaluation

Suite: Kahn1 held-out set (14,663 items: banking intent, sentiment, NLI; the maker's own split). Numbers are the publisher's own.

- Decision accuracy: 72.4%
- Calibration error (ECE): 0.072
- Median latency: 88.7 ms
- p95 latency: 279 ms

## Model card

<!-- generated by scripts/seed_catalog.py; edit content/models/catalog.yaml -->

# Kahn1-Qwen3.5-4B

Open 4B decision model by Axel Montzamir: reads a text state and answers typed choice, score and noul questions in one forward pass, taking probabilities from the option tokens with permutation debiasing. Apache-2.0 weights; the 3B sibling is research-only.

A personal project. Answers are read from the logits of the option tokens, giving a full distribution and a confidence per answer, with no text generated; permutation debiasing averages over three option orderings. On the maker's own held-out set (banking intent, sentiment and NLI sources) choice reaches 92.3%, score 51.9% exact match and noul 75.5%. The 3B sibling on Qwen2.5-3B is under the Qwen Research License, so it is not commercial and is only linked. Code is MIT, with a site at kahn1.com.

## What it decides

- **choice** — picks one option from a set
- **score** — places the input on an ordered scale
- **noul** — answers a yes/no question with one probability

## At a glance

| | |
|---|---|
| Parameters | 4B |
| Base model | `Qwen/Qwen3.5-4B` |
| Maker | Axel Montzamir (Okura66) |
| Released | 2026-09-25 |
| License | apache-2.0 |
| Reported accuracy | 72.4% |
| Reported latency | 88.7 ms p50 / 279.3 ms p95 on an RTX 5070 Ti with vLLM, three orderings |

## Get the weights

```bash
pip install systemonemodels
systemone pull okura66/kahn1
```

The files are served from the maker's Hugging Face repository, [`Okura66/Kahn1-Qwen3.5-4B`](https://huggingface.co/Okura66/Kahn1-Qwen3.5-4B), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/Okura66/Kahn1-Qwen3.5-4B)
- [Code](https://github.com/Okura66/kahn1)
- [Site](https://kahn1.com)

---

*This page was opened by System One for Axel Montzamir (Okura66), who can claim the organisation and take it over at any time.*

---

From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models