Xor 26B-A4B v1.2

Post-trained Gemma 4 26B-A4B for typed decision tasks, distributed as merged BF16 weights. No adapter merge is required for inference. The bundled wrapper exposes /v1/systemone for text, image, and video input and noul, choice, and score questions.

Base: google/gemma-4-26B-A4B-it, revision 4d7ae4984b7db7de8f8457170b3f1a419ee76d52. Architecture, tokenizer, and processor are unchanged from v1.1.

What changed from v1.1

v1.2 is a weight blend of two post-trained checkpoints of the same base:

• Non-expert tensors: 0.60 × the v1.1 checkpoint (xn-lora-xl-5050) + 0.40 × a second post-trained checkpoint (ndnew, a LoRA merged onto the base and trained on a separate decision-task pool), computed in float and cast back to BF16.
• Routed-expert tensors (experts.down_proj and experts.gate_up_proj in all 30 layers, 60 tensors): taken unchanged from the v1.1 checkpoint.

The 0.60/0.40 split was chosen after comparing several blends on public Decision Index results, so any advantage over simpler choices is partly tuned on that public suite.

The serving bundle is the v1.1 bundle with one configuration change: the temperature map is now choice=1.95, noul=1.0, score=1.0 (v1.1: choice=2.0, noul=1.7, score=1.0). The wrapper source and SGLang image are byte-identical to v1.1. The choice temperature was carried over from an earlier checkpoint's fit and was not re-fitted on held-out data for this blend.

Revisions

main contains v1.2; the v1.2 tag pins this release. The v1 and v1.1 tags and historical revisions remain available. Pin the tag for reproducible inference. The NVFP4 companion repository is a separate release and is not updated by this BF16 release.

Public evaluation

Decision Index 0.3, self-run on 2026-10-16 with this repository's serving bundle, using upstream commit f5a6e4fd85fed05e77878dfc4966d0568bfb45ec:

Hardware: two NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs (96 GB each), one isolated TP1 replica per GPU started with this bundle's run.sh, six simultaneous requests per replica. Twelve unchanged upstream sequential runners processed disjoint shards, followed by unchanged upstream scoring. Latencies above include wrapper and inference under load; they are not single-GPU sequential eligibility measurements. The result is the self-run public index, not the private Full score or a maintainer-verified leaderboard result. Aggregate reports are in evaluation/.

Reproducible serving

Requires a Linux NVIDIA GPU host, Docker Compose with GPU support, NVIDIA Container Toolkit, and approximately 96 GB GPU memory for the tested settings. CPU-only and smaller-GPU configurations are not validated. The public model download requires the Hugging Face CLI.

hf download juspay/Xor-26B-A4B --revision v1.2 --local-dir ./xor-v1.2
export MODEL_DIR="$(pwd)/xor-v1.2"
mkdir -p ./xor-v1.2-runtime
tar -xzf "$MODEL_DIR/serving/xor-v1.2-serving.tar.gz" -C ./xor-v1.2-runtime
cd ./xor-v1.2-runtime
export GPU_ID=0
export API_PORT=8000
bash ./run.sh

The API binds to localhost, has no API-key authentication, and should not be exposed directly to the Internet. Use an authenticated gateway when external access is needed.

curl -fsS http://127.0.0.1:8000/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{"model":"xor-26b-a4b-v1.2","state":"The customer was charged twice.","questions":{"billing":{"type":"noul","instructions":"Is this a billing issue?"}}}'

Stop this Compose project with bash ./stop.sh. It does not delete model weights.

Inference settings and limits

Pinned SGLang image: prakhar1611/xor-sglang@sha256:94c48d2a6cc98dc456cf93f723707ea7dd81dddfe1061e823b348d68bbe8158f, built from lmsysorg/sglang@sha256:6bcaa47db52f78ce0d67863b8b2431221b79bc23204a80cad757fa819d00e921 with the tensor/list candidate-logprob handling fix. The exact frozen wrapper is bundled; two-order scoring is enabled, with temperatures choice=1.95, noul=1.0, and score=1.0.

The wrapper scores candidate label-token probabilities rather than generating explanatory prose. choice supports 2–255 options; score supports 2–10 ordered levels. Context limit is 131,072 tokens. Confidence is the largest candidate probability. The wrapper does not support PDF files, audio, arbitrary text extraction, or general chat generation.

Requests may include images, an array of 1–8 base64 image data URLs, or video, one base64 video data URL. Send only one kind of media per request. Media is processed with Gemma's own chat template and the matching processor embedded in processor_config.json; all questions and reverse-order scoring rows receive the same supplied media. Video is processed as sampled frames, without its audio track. Request bodies larger than 8 MB are rejected. Only data URLs are accepted; the wrapper does not fetch client-supplied remote URLs. Expanded media tokens consume context capacity and are subject to backend context checks. Image and video checks for this release are functional smoke tests, not a multimodal benchmark.

Public benchmark evaluation is not a guarantee of downstream quality or calibrated probabilities.

Training data and known gaps

No claim of benchmark-clean training is made. Known issues:

• The v1.1 checkpoint's training mixes contain test rows from the public suite's POP909 benchmark. v1.2 inherits this through the v1.1 tensors.
• The second checkpoint's training pool was screened against the Decision Index 0.3 public suite. Whole-field screening found no matches; a stricter per-field screen found 26 exact matches and about 3,500 near-matching contexts, almost all shared tool-retrieval contexts.
• The second checkpoint's training rows do not carry per-row source or licence provenance.

License

Apache License 2.0. See LICENSE and THIRD_PARTY_NOTICES.md for notices. Benchmark inputs are not redistributed in the model repository; their upstream dataset terms continue to apply.