Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

JevK5

Apache-2.0 open-weight Jev alternative that answers yes/no, choice and score questions with per-option probabilities in one pass.

LibrarybetaApache-2.0
View on GitHub ↗

Overview

JevK5 is an independent open-weight decision model built on Qwen3.5 (4B, plus a 9B variant) with a distilled LoRA merged in. It reads a state such as a ticket, log, policy or diff and answers noul, choice and score questions by taking a softmax over the answer letters' next-token logits, so it generates no tokens. A server exposes a TypeSafe-style /v1/systemone endpoint, so existing Jev clients can point at it, and GGUF builds are listed for llama.cpp.

It is aimed at teams that want Jev-style typed decisions on their own GPU. The README reports about 13 ms per short decision on an H100 using CUDA graphs, and cites JevBench v1.4 ranking v0.2 second of 76 systems (first among open entrants) and the Jev Decision Index placing it 15th of 49. It is English-only, handles up to 16 options per pass and refuses inputs over 16,384 tokens. The project is not affiliated with TypeSafe.

What the decision model decides

TypeQuestionAnswers
Yes / noAny yes/no question about the supplied state—
ChoiceWhich of up to 16 options fits the state (more via multiple passes)?—
ScoreWhere does the state sit on an ordered rubric?—

How it works

  1. Send a state and typed questions to the /v1/systemone endpoint
  2. Each question is formatted with lettered options and run through the model once
  3. Option probabilities come from a temperature-scaled softmax over the answer-letter logits
  4. Client thresholds or branches on the returned probabilities

Strengths

  • Open weights and code under Apache-2.0, self-hostable
  • Reported ~13 ms per short decision on an H100
  • Wire-compatible with TypeSafe's request shape
  • Independent benchmark results are linked rather than self-reported only

Caveats

  • Needs a GPU (about 9 GB bf16 for 4B, 19 GB for 9B)
  • English-only and not yet measured on Jev's published real-world workflows
  • Requests are serialised on one GPU

Getting started

Weights are on Hugging Face at alibiserikbay/JevK5; the repository's server exposes /v1/systemone for Jev clients.

At a glance

Type
Library
Status
beta
Language
Python
Runtime
GPU inference server (CUDA), GGUF for llama.cpp
Has code
yes
License
Apache-2.0
Stars
150

Use cases and integrations

How curated lists describe it

  • Open alternative: an open-weight model answering yes/no, choice and score questions with a probability per option in one forward pass, reporting about 13 ms on an H100 and 33.1% against Jev's 36.7% on 308 sealed decisions.

    yibie/awesome-jev

  • Apache-2.0 decision models on Qwen3.5 (2B, 4B, 9B) behind a `/v1/systemone` server, with GGUF builds for llama.cpp on CPUs and GPUs

    AnotiaWang/awesome-decision-models

  • An open model that answers the same typed questions Jev answers — yes/no, choice, score — in one forward pass with zero generated tokens (~13 ms on an H100), plus a head-to-head harness that puts Jev and the open model on the identical board and question so the two can be compared directly. It serves TypeSafe's \`/v1/systemone\` wire format, so a Jev client can point at it unchanged.

    logicrw/awesome-jev-projects

  • JevK5: open-weight alternative to TypeSafe Jev. Typed decisions with probabilities in one forward pass; Apache-2.0 weights and code.

    wh000wh000/awesome-jev-live

  • JevK5: open-weight alternative to TypeSafe Jev. Typed decisions with probabilities in one forward pass; Apache-2.0 weights and code.

    hellogumbo/awesome-jev

  • Open-weight decision model that serves the `/v1/systemone` request shape from your own GPU, with weights on Hugging Face.

    MrJev/awesome-jev

Frequently asked questions

How does JevK5 differ from Jev?

It uses open Qwen3.5 weights with an option-logit readout rather than Jev's unpublished architecture, and you host it yourself.

Can I use my existing Jev client?

Yes, the server accepts the TypeSafe-style /v1/systemone request shape.

How good is it?

The README cites JevBench v1.4 ranking v0.2 second of 76 systems; on its 308 sealed decisions JevK5 scored 33.1% against Jev's 36.7%.

Related projects

  • TypeAR★ 946

    TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.

    ProjectbetaPython
  • jev-dllm

    Reads Jev-style choice, yes/no and score probabilities from masked diffusion language models with shared Yes/No weights.

    ProjectresearchPython
  • metask-jev

    Open Jev-class decision models on Qwen3.5 with candidate-logit readout; the 4B model reports 80.1% on JevBench's 231 items.

    ProjectresearchPython
  • system-one-open

    Open Jev replica on Gemma 4 E2B and Gemma 3 270M served on Modal, with eight demos and a TypeSafe eval head-to-head.

    ProjectresearchPython
  • jev-position-test

    Small experiment: reordering enum options changed a local Jev clone's answers on its old default, while hosted Jev's did not.

    DemoresearchPython
  • lev

    Open System One decision model (LoRA on Qwen3.5-4B) that reads typed answers from logits in one pass, with its benchmark harness.

    ModelbetaPython