# Foxl AI: bobcat

> Typed-decision model from Foxl AI: Qwen3.8-27B with a merged LoRA that returns a probability for every option you name (choice, noul, score) from one forward pass, with no generated text. Uses TypeSafe's System One request shapes. Apache-2.0; English and Korean.

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

## Facts

| | |
|---|---|
| Maker | Foxl AI (https://systemonemodels.ai/foxl-ai) |
| Decides | choice, score, noul |
| Architecture | bobcat |
| Base model | qwen/qwen3.8-27b |
| Parameters | 28B |
| Context | 16K tokens |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-09-27 |
| Latest version | 1.1.0 |

## Reported evaluation

Suite: Bobcat sealed final: four Korean tasks, 1,614 decisions from KLUE MRC and Wizard of Seoul, opened once (Foxl AI's own). Numbers are the publisher's own.

- Decision accuracy: 94.3%
- Calibration error (ECE): 0.012
- Median latency: 42.8 ms

## Model card

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

# Bobcat 1.1

Typed-decision model from Foxl AI: Qwen3.8-27B with a merged LoRA that returns a probability for every option you name (choice, noul, score) from one forward pass, with no generated text. Uses TypeSafe's System One request shapes. Apache-2.0; English and Korean.

The weights are published by Sanghwa Na on his personal Hugging Face account; the code is in the foxl-ai GitHub organisation. The typed contract needs the maker's bobcat.api_server, because plain `vllm serve` only exposes text generation. Requests follow TypeSafe's published System One shape; Foxl says it is not affiliated with TypeSafe and that no outputs of Jev or any other teacher were used in training. The sealed final and the development split are Korean; English evidence rests on a handful of public examples. Calibration is one temperature (1.2008), and the maker says confidence is not the probability of being right. Weak on insufficient evidence, on arithmetic and dates; 7.8% of answers still move under injected wrong answers. Needs a large GPU (tested on an RTX PRO 6000 96 GB). Siblings: an NVFP4 build for Blackwell and Bobcat Flash 1.1 (Gemma 4 26B-A4B, about 4B active, distilled from the 27B). Text and JSON inputs only.

## 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 | 27.8B |
| Base model | `Qwen/Qwen3.8-27B` |
| Maker | Foxl AI |
| Released | 2026-09-27 |
| License | apache-2.0 |
| Reported accuracy | 94.3% |
| Reported latency | 42.8 ms p50 per decision (512 tokens, 8 candidates), NVFP4 on one RTX PRO 6000 Blackwell, vLLM 0.30.0, in-engine without HTTP |

## Get the weights

```bash
pip install systemonemodels
systemone pull foxl-ai/bobcat
```

The files are served from the maker's Hugging Face repository, [`sanghwa-na/bobcat-1.1`](https://huggingface.co/sanghwa-na/bobcat-1.1), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/sanghwa-na/bobcat-1.1)
- [NVFP4 build](https://huggingface.co/sanghwa-na/bobcat-1.1-nvfp4)
- [Bobcat Flash 1.1](https://huggingface.co/sanghwa-na/bobcat-flash-1.1)
- [Code](https://github.com/foxl-ai/bobcat)
- [Technical write-up](https://foxl.ai/blog/bobcat-typed-decisions)

---

*This page was opened by System One for Foxl AI, 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