XOR Lite
XOR Lite (xor-lite) is a post-trained version of openbmb/MiniCPM5-2B for typed decision tasks. It is served through a TypeSafe-compatible /v1/systemone API
Release
| Field | Value |
|---|---|
| Base checkpoint | openbmb/MiniCPM5-2B |
| Base revision | 12a3808a956f869c767195e9266b59c4d21d92e2 |
| Architecture | Dense LlamaForCausalLM |
| Parameters | Approximately 2.52 billion |
| Released precision | BF16 |
| License | Apache License 2.0 |
| Packaging | Fully merged weights; no adapter loading is required |
Pin the published revision for reproducible use
hf download juspay/xor-lite --revision REPLACE_WITH_RELEASE_REVISION --local-dir xor-lite
Interface
The server accepts a state and a map of typed questions
noul: binary probabilitychoice: categorical decision and full probability distributionscore: expected ordinal score and full probability distribution
The serving layer performs constrained single-token candidate readout, forward and reverse option-order evaluation, probability calibration, and SystemOne schema conversion. It supports up to 255 candidates using the included MiniCPM-specific marker table. The released serving layer is part of the inference configuration required for reproducible results
Quick start
hf download juspay/xor-lite --revision REPLACE_WITH_RELEASE_REVISION --local-dir xor-lite
(cd xor-lite/serving && sha256sum -c xor-lite-serving.tar.gz.sha256)
mkdir -p xor-lite-runtime
tar -xzf xor-lite/serving/xor-lite-serving.tar.gz -C xor-lite-runtime --strip-components=1
cd xor-lite-runtime
cp .env.example .env
sed -i "s|^MODEL_DIR=.*|MODEL_DIR=$(cd ../xor-lite && pwd)|" .env
./run.sh
When the smoke test succeeds, the API is available at http://127.0.0.1:30002/v1/systemone
Public JEVBench self-run
XOR Lite was evaluated locally on all 231 public JEVBench decisions using harness commit fd54ea7dc02bbe29c6ac8f6e015a54cdcff26805, the existing typesafe adapter, and one request at a time
The run used the released configuration on 1 x NVIDIA H200 NVL (143 GB), tensor parallelism 1, data parallelism 1, with request caching disabled
| Tier | Attempted | Valid | Correct | Accuracy | Brier | ECE | p50 | p95 |
|---|---|---|---|---|---|---|---|---|
| Easy | 48 | 48 | 48 | 1.0000 | 0.1467 | 0.2294 | 0.0250 s | 0.0279 s |
| Original | 72 | 72 | 66 | 0.9167 | 0.2777 | 0.2986 | 0.0251 s | 0.0291 s |
| Hard public | 111 | 111 | 55 | 0.4955 | 0.6153 | 0.0902 | 0.0308 s | 0.0659 s |
| All public | 231 | 231 | 169 | 0.7316 |
Operational success, coverage, schema validity, and strict schema validity were 1.0000 for each tier. These are self-run public-tier results, not an official JEVBench rank. Latency is hardware-specific and was measured locally without network overhead
Validated runtime
| Setting | Value |
|---|---|
| SGLang image | lmsysorg/sglang@sha256:6bcaa47db52f78ce0d67863b8b2431221b79bc23204a80cad757fa819d00e921 |
| Tensor parallelism | 1 |
| Data parallelism | 1 |
| Validated GPU | 1 x NVIDIA H200 NVL, 143 GB |
| Maximum prefill tokens | 250,000 |
| Static memory fraction | 0.50 |
| Probability temperature | 100.0 for noul, choice, and score |
NVIDIA L4 validation
The immutable v1.0 release was also validated without modification on one NVIDIA L4 with 24 GB of memory using tensor parallelism 1 and data parallelism 1. The complete public JEVBench run reproduced the H200 result exactly at 169/231 correct, with no operational, coverage, or schema failures
| Tier | Correct | Accuracy | p50 | p95 |
|---|---|---|---|---|
| Easy | 48/48 | 1.0000 | 0.0458 s | 0.0481 s |
| Original | 66/72 | 0.9167 | 0.0465 s | 0.0488 s |
| Hard public | 55/111 | 0.4955 | 0.1175 s | 0.5868 s |
| All public | 169/231 | 0.7316 |
The loaded service used 13,803 MiB of the L4's 23,034 MiB. The run used NVIDIA driver 595.91.07 and the same pinned SGLang image shown above. The evidence archive has SHA-256 6389b280c9abfffd608a3864a1b8ed7570974ca415ad5f079d0224f429694ed9
The L4 result demonstrates that the H200 used for the original release measurement is not a minimum hardware requirement. It does not define or announce a hosted-service tariff
The service binds to localhost by default and does not require an API key. Remote deployments must add authentication, TLS, rate limits, and request-size limits at the ingress layer