XOR Omni
XOR Omni (xor-omni) is a multimodal decision model post-trained from google/gemma-4-26B-A4B-it. It is designed for typed decision tasks and structured one-pass readout.
Model details
| Field | Value |
|---|---|
| Base model | google/gemma-4-26B-A4B-it |
| Architecture | Gemma 4 multimodal mixture-of-experts |
| Parameters | Approximately 26B total, about 4B active per token |
| Released precision | BF16 |
| Context length | 32,768 tokens |
| Input modalities | Text and image |
| Output | Text |
| Runtime | SGLang |
| License | Apache License 2.0 |
| Packaging | Fully merged weights; no adapter loading is required |
Intended use
XOR Omni is intended for typed decision workloads, including:
- Binary decisions
- Categorical choices
- Ordinal scoring
- Structured readout from combined text and image context
It is optimized for deterministic, one-pass decision output rather than open-ended assistant conversation.
Quick start
Download the model:
hf download juspay/xor-omni --local-dir xor-omni
Example SGLang launch:
python -m sglang.launch_server \
--model-path ./xor-omni \
--trust-remote-code \
--context-length 32768 \
--host 0.0.0.0 \
--port 30000
Use the included chat_template.jinja when serving the model. The exact deployment configuration should be validated for your target hardware and workload.
Limitations
- This model is specialized for typed decision tasks and is not intended as a general-purpose assistant.
- Multimodal behavior should be validated on your own data before production use.
- Performance may vary with prompt format, context length, hardware, and serving configuration.
- This repository contains merged model weights only; training scripts, adapters, benchmark artifacts, and logs are intentionally not included.
License
This model is released under the Apache License 2.0, inherited from the base Gemma 4 checkpoint.