GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
juspay
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
juspay
Model Tree
google/gemma-4-26B-A4B-it
Input Modalities
Output Modalities
Supported Functionality
main contains v1.1; the v1.1 tag pins this release. The existing v1 tag 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.
Decision Index 0.3, self-run on 2026-10-09 using upstream commit 9eb2dbe2a358004c8782c66e40a83ac07b953fec:
| Metric | Result |
|---|---|
| Public Decision Index | 57.62 |
| Requests attempted | 140,620 |
| Requests scored after upstream exclusions | 140,178 |
| Errors | 0 |
| Coverage | 100% |
| Mean HTTP request latency | 196.4 ms |
| Median HTTP request latency | 137.0 ms |
| p95 HTTP request latency | 461.6 ms |
Hardware: six NVIDIA B300 SXM6 AC GPUs, 275040 MiB each. Six isolated TP1 replicas, four simultaneous requests per replica, 24 total. These latencies include wrapper and inference under load; they are not single-GPU sequential eligibility measurements. Twenty-four unchanged upstream sequential runners processed disjoint shards, followed by unchanged upstream scoring. Fresh uninterrupted run, no resume or reused responses; all workers exited zero. The 57.62 is the self-run public index, not the private Full score or a maintainer-verified leaderboard result. Aggregate reports are in evaluation/.
That Decision Index run used the released multimodal wrapper identified by SHA-256 2eed569347cd5c6d4afbc638c0feb1b3cdbbb4dff02ed927788d9ded8cf4e435 and the patched image pinned below. A paired public JEVBench regression on one RTX PRO 6000 GPU returned 48/48 Easy, 69/72 Original, and 87/111 Hard for both the original text-only and extended wrappers, with zero failures. Small probability differences between runs occurred; identical floating-point outputs are not guaranteed by the backend. Image/video smoke tests are functional checks, not a multimodal benchmark.
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.
bash
hf download juspay/Xor-26B-A4B --revision v1.1 --local-dir ./xor-v1.1export MODEL_DIR="$(pwd)/xor-v1.1"mkdir -p ./xor-v1.1-runtimetar -xzf "$MODEL_DIR/serving/xor-v1.1-serving.tar.gz" -C ./xor-v1.1-runtimecd ./xor-v1.1-runtimeexport GPU_ID=0export API_PORT=8000bash ./run.shThe 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.
bash
curl -fsS http://127.0.0.1:8000/v1/systemone \ -H 'Content-Type: application/json' \ -d '{"model":"xor-26b-a4b-xl5050","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.
Pinned SGLang image: prakhar1611/xor-sglang@sha256:94c48d2a6cc98dc456cf93f723707ea7dd81dddfe1061e823b348d68bbe8158f. It contains the tensor/list candidate-logprob handling fix. The exact frozen wrapper is bundled; two-order scoring is enabled, with temperatures choice=2.0, noul=1.7, 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.
Public benchmark evaluation is not a guarantee of downstream quality or calibrated probabilities. No claim of benchmark-clean training is made. Training overlap has not been independently established for this checkpoint.
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.
hf download juspay/Xor-26B-A4B --revision v1.1 --local-dir ./xor-v1.1export MODEL_DIR="$(pwd)/xor-v1.1"mkdir -p ./xor-v1.1-runtimetar -xzf "$MODEL_DIR/serving/xor-v1.1-serving.tar.gz" -C ./xor-v1.1-runtimecd ./xor-v1.1-runtimeexport GPU_ID=0export API_PORT=8000bash ./run.shcurl -fsS http://127.0.0.1:8000/v1/systemone \ -H 'Content-Type: application/json' \ -d '{"model":"xor-26b-a4b-xl5050","state":"The customer was charged twice.","questions":{"billing":{"type":"noul","instructions":"Is this a billing issue?"}}}'