The vLLM project released version 0.32.0 on October 18, 2026, adding /v1/systemone as a default-enabled endpoint for structured decisions on supported models. Requests supply a state and typed questions, and responses include answer probabilities. Sources: vLLM, ModelSystem.One
The choice type selects an option, noul returns a yes probability, and score returns an expected level indexed from zero. The documented limits are 64 questions per request, 26 choice options and 10 score levels. Sources: vLLM
The generic generation path calculates probabilities from first-token label log-probabilities, normalized across allowed labels. Each label must tokenize as a single token. Supported generation architectures require --logprobs-mode set to raw_logprobs or processed_logprobs. Unsupported read strategies return HTTP 501. Clef-Omni uses a pooling model to score all questions’ options in one forward pass. Sources: vLLM
Identical requests can return different probabilities depending on batching and prefix-cache state, according to the documentation. For repeatability, it prescribes VLLM_BATCH_INVARIANT=1, which reduces throughput. ModelSystem.One corroborated the endpoint’s release and behavior through upstream materials but did not report independent runtime testing. Sources: vLLM, ModelSystem.One
The release also removes Python 3.10 support and restricts default origins for cross-origin resource sharing. Its dependencies include transformers>=5.16.1,<5.20.0 and tokenizers>=0.23.2. Sources: vLLM