ENPT

openjev-sglang

Last updated:

RuntimeMIT

⚠️ Why not a model

TypeSafe-compatible endpoint over Qwen via SGLang prefill/logprob hack — not a trained decision model.

Technical specs

Base LLMs
Qwen3.6-35B-A3B
Decision types
choice, score, boolean
Features
sglang, prefill-logprob, typesafe-compatible
License
MIT

Links

openjev-sglang provides a TypeSafe-compatible /v1/systemone endpoint over Qwen3.6-35B-A3B using SGLang’s prefill and logprob capabilities.

How it works

The runtime uses SGLang’s efficient inference features:

  1. Prefill the context and question
  2. Extract logprobs for each decision option
  3. Return structured decisions in TypeSafe format

This is an inference hack — not a dedicated decision model.

Why it’s not a model

openjev-sglang is infrastructure, not a model:

  • Uses frozen Qwen3.6-35B-A3B weights
  • No training or fine-tuning involved
  • Decision probabilities come from logprob extraction

Key features

  • TypeSafe-compatible API: Drop-in replacement for /v1/systemone
  • SGLang backend: Efficient KV caching and prefill
  • Large model support: Uses Qwen3.6-35B-A3B (MoE)