vLLM v0.32.0 serves /v1/systemone by default, and transformers serve now has the endpoint too
publicado (OK do Fabricio 18/10 BRT).
Two mainstream inference stacks now speak the /v1/systemone shape that TypeSafe’s Jev introduced.
vLLM v0.32.0 was published on 18 October 2026 (03:50 BRT). Its release notes list the /v1/systemone endpoint (new in #59299) as “served by default” (#60677, #60696). The docs page on structured decisions models says it answers choice, noul and score questions by reading the log-probabilities of the first generated token, so the server has to run with --logprobs-mode raw_logprobs or processed_logprobs. A model without a known way to read its answers gets an HTTP 501. The docs list limits of 64 questions per request, 26 options and 10 score levels, and warn that probabilities can shift with batching and prefix-cache state. The same release adds a text-only, OpenAI-style /v1/decisions route.
The documented models include Cloudflare’s Clef-Omni (#61064), Clef-Flash, the vllm-sr d3 family, Qwen3 and Qwen3.5, Liquid’s d1-3B (#60676) and DiffusionGemma (#60466). v0.32.0 pins transformers>=5.16.1,<5.20.0, so it does not yet run with the transformers release below.
transformers v5.20.0 was published on 15 October. PR #49322 (Marc Sun, merged 12 October) adds /v1/systemone to transformers serve for autoregressive models, with an optional --decision-config file. The v5.20.0 release notes do not mention the PR; we checked the commit history, and the release comes 74 commits after the merge, so it includes the feature. An open issue (13 October) reports HTTP 500 errors when chat_template_kwargs are overridden and for models with no chat template.
Why it matters. Until now, serving a decision model meant a vendor API, Ollama on GGUF files, or a smaller project such as vllm-jev. With the endpoint in vLLM and transformers, any team already running those stacks can try open decision models without a new server. The weights still matter more than the server: both stacks only read answers out of a model, and neither trains one. This is infrastructure, so it stays off the quadrant and changes no model scores.
Runtime pages: vLLM and transformers serve. The third-party vllm-jev is a different project and still serves checkpoints that the upstream endpoint does not.