juspay/jev-one

JevOne

JevOne (jev-one) is a post-trained version of Qwen/Qwen3.6-35B-A3B for typed decision tasks. It is served through a TypeSafe-compatible /v1/systemone API

Base model

Field	Value
Base checkpoint	Qwen/Qwen3.6-35B-A3B
Architecture	Mixture-of-experts causal language model
Total parameters	35 billion
Activated parameters	Approximately 3 billion per token
Released precision	BF16
Base license	Apache License 2.0
Packaging	Fully merged weights; no adapter loading or merging is required

Interface

The server accepts a state and a map of typed questions:

• noul: binary probability
• choice: categorical decision and full probability distribution
• score: expected ordinal score and full probability distribution

The serving layer performs deterministic single-token candidate readout, forward and reverse option-order evaluation, probability calibration, and schema conversion. The serving layer is part of the released inference configuration and must be used for reproducible results

Quick start

The following commands download the validated release, verify and extract the serving bundle, and start JevOne on two GPUs:

export JEVONE_REVISION=4de0db772d718134bd32fe5d4c8da77d97a36240
export INSTALL_ROOT="$PWD/jev-one-eval"
export MODEL_DIR="$INSTALL_ROOT/model"

hf download juspay/jev-one   --revision "$JEVONE_REVISION"   --local-dir "$MODEL_DIR"

(
  cd "$MODEL_DIR/serving"
  sha256sum -c jev-one-serving.tar.gz.sha256
)

tar -tzf "$MODEL_DIR/serving/jev-one-serving.tar.gz"

mkdir -p "$INSTALL_ROOT/runtime"
tar -xzf "$MODEL_DIR/serving/jev-one-serving.tar.gz"   -C "$INSTALL_ROOT/runtime"   --strip-components=1

cd "$INSTALL_ROOT/runtime"
export CUDA_VISIBLE_DEVICES=0,1
export TP_SIZE=2
export API_PORT=49001
./run.sh

The setup requires Linux x86-64, the Hugging Face CLI, Docker Engine with Docker Compose v2, the NVIDIA Container Toolkit, and approximately 120 GB of free disk space

When the smoke test succeeds, the API is available at http://127.0.0.1:49001/v1/systemone

Test it with the included request:

curl -sS -X POST http://127.0.0.1:49001/v1/systemone   -H 'Content-Type: application/json'   --data @examples/request.json

Public JEVBench self-run

JevOne was evaluated locally on the public tiers of JEVBench using harness commit fd51755eb0c0b546ca206d764faf3302feca913e, the existing typesafe adapter, and one request at a time

The run used 2 x NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs with 96 GB each, tensor parallelism 2, and the pinned serving image listed below

Tier	Attempted	Valid	Correct	Accuracy	Macro accuracy	Brier mean	ECE	p50	p95
Easy	48	48	48	1.0000	1.0000	0.0018	0.0257	0.0766 s	0.0888 s
Original	72	72	70	0.9722	0.9722	0.0897	0.1285	0.0773 s	0.0812 s
Hard public	111	111	86	0.7748	0.8033	0.3460	0.0579	0.1361 s	0.3887 s

Operational success, coverage, schema validity, and strict schema validity were 1.0000 for all three public tiers

These are self-run public-tier results, not an official JEVBench rank. An independent full-suite evaluation, including held-out items, has been requested in JEVBench issue #20

Validated runtime

Setting	Value
SGLang image	lmsysorg/sglang@sha256:6bcaa47db52f78ce0d67863b8b2431221b79bc23204a80cad757fa819d00e921
Tensor parallelism	2
Validated GPUs	2 x NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB each
Maximum prefill tokens	250,000
Static memory fraction	0.85

For directly comparable latency measurements, use the same GPU model, GPU count, tensor-parallel configuration, and pinned serving image. Latency from other hardware should be identified as hardware-specific

Operational notes

The model files occupy approximately 66 GB. The validated SGLang configuration reserves substantial GPU memory for its KV cache. Smaller prefill budgets may work on other two-GPU configurations but are not covered by the validated setup

The service does not require an API key when bound to localhost. Remote deployments must add authentication, TLS, rate limits, and request-size limits at the ingress layer