Xor, an openDecision Modelyou run yourself

Apache 2.0open weights
35B / 3BMoE, active
1 GPUvalidated

Typed questions in, typed answers out. Xor is post-trained on Qwen3.6-35B-A3B for classification and decisions over text, images and video, and ships with the serving layer that makes its answers calibrated and reproducible.

Try XOR yourself Quick start

typed questions
Ask noul, choice or score questions and get typed answers back, never free text.
calibrated
Every answer comes with a full probability distribution, calibrated per question type.
multimodal
Classify from text plus up to eight images, or one video, in a single request.
order-debiased
Options are read forward and in reverse, so candidate order does not sway the answer.
up to 255 options
Choice and score questions take anywhere from 2 to 255 candidates.
one gpu
Validated on a single NVIDIA H200 with a pinned, patched SGLang runtime.
verifiable release
Every model file is checked against published hashes before the server starts.
self-hosted
Runs on your own infrastructure. No API key needed when bound to localhost.

Try it yourself

Four steps from download to a running API on one GPU.

Linux x86-64
Hugging Face CLI
Docker Engine + Compose v2
NVIDIA Container Toolkit
~120 GB free disk

01Download a pinned release

hf download juspay/xor --revision v1.1 --local-dir xor-1.1

02Verify and extract the serving bundle

(cd xor-1.1/serving && sha256sum -c xor-1.1-serving.tar.gz.sha256)

mkdir -p xor-1.1-runtime
tar -xzf xor-1.1/serving/xor-1.1-serving.tar.gz \
  -C xor-1.1-runtime --strip-components=1

03Point it at the weights and start it

cd xor-1.1-runtime
cp .env.example .env
sed -i "s|^MODEL_DIR=.*|MODEL_DIR=$(cd ../xor-1.1 && pwd)|" .env
./run.sh

04Ask it a typed question

curl -sS -X POST http://127.0.0.1:30002/v1/systemone \
  -H 'Content-Type: application/json' \
  --data @examples/request.json

The API comes up at 127.0.0.1:30002/v1/systemone. Before exposing it beyond localhost, add authentication, TLS, rate limits and request-size limits at your ingress.

Benchmarks

Xor 1.1 on the public tiers of JevBench, run locally on one H200, one request at a time. A self-run result, not an official rank.

Tier	Correct	Accuracy	p50	p95
Easy	48/48	100%	0.074 s	0.077 s
Original	70/72	97.2%	0.074 s	0.078 s
Hard	89/111	80.2%	0.078 s	0.138 s
All public	207/231	89.6%	0.075 s	0.121 s

90.6%macro accuracy
0.170brier mean
0.056ece

Model details

Base checkpoint: Qwen/Qwen3.6-35B-A3B
Architecture: Mixture-of-experts causal LM
Parameters: 35B total, ~3B active per token
Post-training: LoRA rank 16, merged into BF16 weights
Precision: BF16
Max prefill: 250,000 tokens
Model files: ~66 GB
Request limit: 8 MB, ~6 MB of media
License: Apache 2.0