ENPT

vLLM (upstream)

Last updated:

RuntimeApache-2.0~94,000 ★

⚠️ Why not a model

Inference engine. Since v0.32.0 it serves /v1/systemone for models trained by others; it ships no decision weights and reads answers from first-token log-probabilities.

Technical specs

Base LLMs
Clef-Omni, Clef-Flash, vllm-sr d3 family, Qwen3, Qwen3.5, LiquidAI/d1-3B, DiffusionGemma
Decision types
choice, noul, score
Features
systemone-endpoint, served-by-default, first-token-logprobs, multi-question, openai-decisions-route, gpu-serving
License
Apache-2.0

Links

publicado (OK do Fabricio 18/10 BRT).

vLLM, the widely used GPU serving engine, added a /v1/systemone endpoint that is served by default in v0.32.0 (18 Oct 2026). This is the upstream project. It is not vllm-jev, a separate third-party wrapper.

Supported surface

  • Endpoint: POST /v1/systemone (new in #59299, on by default in #60677 and #60696). A text-only OpenAI-style /v1/decisions route comes in the same release.
  • Question types: choice, noul and score. The docs give limits of 64 questions per request, 26 options and 10 score levels, and say a label is the first token of the reply.
  • Read from log-probabilities: start the server with --logprobs-mode raw_logprobs or processed_logprobs. A model without a known read strategy gets HTTP 501.
  • Documented models: Clef-Omni (#61064), Clef-Flash, the vllm-sr d3 family, Qwen3 and Qwen3.5, Liquid d1-3B (#60676) and DiffusionGemma (#60466).

Caveats

  • The docs say probabilities can vary with batching and prefix-cache state, so do not treat them as bit-for-bit repeatable.
  • v0.32.0 pins transformers>=5.16.1,<5.20.0, so it does not yet pair with transformers v5.20.0, which has its own transformers serve endpoint.
  • Answer quality depends entirely on the model. Calibration of a base Qwen model read this way is not the same as a model trained to give calibrated probabilities.

Why it is not a model

vLLM trains nothing. It is serving infrastructure that reads typed answers out of models other teams trained, in the same class as Ollama, Ollaya and vllm-jev, so it stays off the quadrant.