jevbooks

← All projects

openjev-sglang

ekzhang/openjev-sglang · Homepage

Jev-compatible HTTP API served by Qwen3.6-35B-A3B on SGLang.

What it is

A server implementing the TypeSafe/Jev HTTP API with Qwen3.6-35B-A3B on SGLang, using prefill-only one-token inference. It targets developers who want a self-hosted, Jev-compatible evaluation endpoint for Noul, Choice, and Score questions.

How it uses Jev

Jev questions (Noul, Choice, Score) are sent to POST /v1/systemone with a state and questions. The server renders the chat template once, warms SGLang's radix cache with the common prefix, then sends each question suffix with max_new_tokens=1 and reads answer-label logprobs, renormalized with softmax.

Primitives:choicescorenoul

Technique worth stealing

Prefill-only one-token readout: request token_ids_logprob for answer labels with max_new_tokens=1, no autoregressive continuation.

Try it

uv sync; uv run modal run modal_app.py, or uv run modal deploy modal_app.py for a public endpoint.

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricunclear0.54
  • Shows a System One patternno0.06
  • Handles uncertaintyno0.03
  • Measuredno0.05
  • Runnableyes0.98
  • Worth recommendingno0.35
  • Model replicayes0.83
  • Problem scopescore on a 0–2 scale1.00
  • About Jevyes0.98

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.