transformers serve
⚠️ Why not a model
Reference inference server. Since v5.20.0 `transformers serve` exposes /v1/systemone for any autoregressive model; it has no decision weights of its own.
Technical specs
- Base LLMs
- any autoregressive model
- Decision types
- choice, noul, score
- Features
- systemone-endpoint, decision-config-file, local, python, multi-question
- License
- Apache-2.0
Links
publicado (OK do Fabricio 18/10 BRT).
Hugging Face’s transformers serve now has a /v1/systemone endpoint. PR #49322 (Marc Sun) was merged on 12 Oct 2026 and adds it for autoregressive models, with an optional --decision-config file.
Version note
v5.20.0 was published on 15 Oct 2026. Its release notes do not list the PR. The release is 74 commits after the merge commit, so the feature is in it; we confirmed that from commit order, not from the notes.
Known problems
Issue #49553 (13 Oct, open when we checked on 18 Oct) reports HTTP 500 errors when chat_template_kwargs are overridden and for models without a chat template.
Why it is not a model
transformers serve reads answers out of a model someone else trained. It adds no decision weights, so it stays off the quadrant, next to vLLM, Ollama and vllm-jev. vLLM v0.32.0 pins transformers<5.20.0, so the two upstream endpoints do not yet share a release.