alibi-serikbay/jevk50.3.0

Qwen3.5-4B with a merged rank-16 LoRA distilled from 17,408 teacher questions and 30,000 public training rows, read out as a temperature-scaled softmax over answer-letter logits. Also in 9B, 2B and GGUF.

choicescorenoulclassifyroutesystem-oneqwenloradistilled4b

JevK5

Qwen3.5-4B with a merged rank-16 LoRA distilled from 17,408 teacher questions and 30,000 public training rows, read out as a temperature-scaled softmax over answer-letter logits. Also in 9B, 2B and GGUF.

Version 0.3 scores 78.4% on JevBench's public hard tier (standard 94.4%, easy 100%) with a calibration error of 0.035 on held-out teacher questions; JevBench v1.4 ranked the previous version second of 76 entries. The Hub lists it as text generation, but decisions come from the jevk5 runtime's readout, and jevk5-serve exposes /v1/systemone. Teacher data was generated under OpenAI's terms. Median latency is 13.2 ms on an H100 for short inputs. Siblings: JevK5-9B, JevK5-2B and JevK5-GGUF.

What it decides

  • choice — picks one option from a set
  • score — places the input on an ordered scale
  • noul — answers a yes/no question with one calibrated probability
  • classify — assigns a category from a fixed taxonomy
  • route — sends the input to one of several destinations

At a glance

Parameters4B
Base modelqwen/qwen3.5-4b
MakerAlibi Serikbay
Released2026-09-22
Licenseapache-2.0
Reported accuracy78.4%
Reported latency13.2 ms p50 / 14.6 ms p95 on H100 (short); ~30 ms p50 with 1–4k-token documents

Get the weights

pip install systemonemodels
systemone pull alibi-serikbay/jevk5

The files are served from the maker's Hugging Face repository, alibiserikbay/JevK5, and verified against the checksums recorded here.

Read more


This page was opened by System One for Alibi Serikbay, who can claim the organisation and take it over at any time.