# Juspay: xor

> Juspay's open decision model. Qwen3.6-35B-A3B with a merged rank-16 LoRA, shipped with a pinned SGLang serving bundle that answers Choice, Score and Noul questions over text, images or a video through a TypeSafe-compatible /v1/systemone API.

- Page: https://systemonemodels.tech/juspay/xor
- API: https://api.systemonemodels.tech/v1/models/juspay/xor
- Download: `pip install systemonemodels && systemone pull juspay/xor`

## Facts

| | |
|---|---|
| Maker | Juspay (https://systemonemodels.tech/juspay) |
| Decides | choice, score, noul, classify, route |
| Architecture | xor |
| Base model | qwen/qwen3.6-35b-a3b |
| Parameters | 35B |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-09-21 |
| Latest version | 1.2.0 |

## Reported evaluation

Suite: JevBench public set (231 items), maker's self-run of Xor 1.2. Numbers are the publisher's own.

- Decision accuracy: 90.0%
- Calibration error (ECE): 0.073
- Median latency: 69 ms
- p95 latency: 162 ms

## Model card

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

# Xor 1.2

Juspay's open decision model. Qwen3.6-35B-A3B with a merged rank-16 LoRA, shipped with a pinned SGLang serving bundle that answers Choice, Score and Noul questions over text, images or a video through a TypeSafe-compatible /v1/systemone API.

A mixture of experts with about 3B parameters active per token. The serving layer reads one candidate token per option, scores the options in forward and reverse order, calibrates and converts to the System One schema; Juspay says it is part of the release and needed for reproducible results. Up to eight images or one video per request. Juspay's own run of Xor 1.2 on JevBench's 231 public items: 90.0% (hard tier 82.0%), ECE 0.073. The main branch still points at Xor 1.1; 1.2 lives on the release/xor-1.2 branch. Siblings: Xor Lite (MiniCPM5-2B, 73.2% on the same public items, runs on one L4), Xor Omni (Gemma-4-26B-A4B, text and images, 32K context, weights only) and Xor NVFP4 (39 GB instead of 66 GB). JevBench v1.4.2 ranked the first public build (then called JevOne) 52nd, with 89.6% public and 33.8% sealed accuracy.

## 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 calibrated probability
- **classify** — assigns a category from a fixed taxonomy
- **route** — sends the input to one of several destinations

## At a glance

| | |
|---|---|
| Parameters | 35B |
| Base model | `qwen/qwen3.6-35b-a3b` |
| Maker | Juspay |
| Released | 2026-09-21 |
| License | apache-2.0 |
| Reported accuracy | 90.0% |
| Reported latency | 69 ms p50 / 162 ms p95 on one H200 NVL (TP1), measured locally without network |

## Get the weights

```bash
pip install systemonemodels
systemone pull juspay/xor
```

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

## Read more

- [Weights and serving bundle](https://huggingface.co/juspay/xor)
- [Xor Lite](https://huggingface.co/juspay/xor-lite)
- [Xor Omni](https://huggingface.co/juspay/xor-omni)
- [Xor NVFP4](https://huggingface.co/juspay/xor-nvfp4)

---

*This page was opened by System One for Juspay, who can claim the organisation and take it over at any time.*

---

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