How does Malkuth compare to the hosted Jev API?

Jev is the hosted API with undisclosed weights, while Malkuth-4B and 2B are post-trained from Kev bases with published evaluation tables. On held-out suites Jev leads accuracy while Malkuth-4B leads on the full set including trained-on data.

What is the catch in getting the quick start running?

Clone Kev, sync with the serve extra, and run the serve module with the model name; the digest pins a Kev revision for reproducing scores. Requests carry their own criteria with several questions per input.

What question shapes and answer fields do I code against?

Choice supports up to 255 options, plus yes-or-no Noul and ordered Score levels. Each response includes probabilities with confidence, so you gate actions in code.

Do the millisecond latency numbers apply to my machine?

The latency table covers one A100 in bf16 with median timings plus a faster repeat-input number from prefix cache. Your CPU or GPU numbers will differ; treat the table as a shape guide.

What are the two Malkuth sizes and their base models?

Malkuth-4B is post-trained from Qwen3.5-4B-Base and Malkuth-2B from Qwen3.8-2B-Distill. Both are post-trained from Kev and served through the TypeSafe System One API.

How does Malkuth-4B score against the Jev API?

On the held-out suites Malkuth-4B reaches 0.724 accuracy against Jev's 0.754, and 0.794 against Jev's 0.790 across all 29 suites. Its Brier score is 0.385 against Jev's 0.355.

What license covers the Malkuth weights?

The code is Apache-2.0, but the weights are research use only because the training mix includes XNLI (CC-BY-NC-4.0) and RACE (non-commercial research only).

CommunityJev Demos & Experiments

Malkuth

Small multilingual decision model for classification with a Korean focus that returns probabilities over a fixed choice set, in 4B and 2B sizes served through Kev.

See highlights

Open github.com

Stars: 5
Language: Python
License: APACHE-2.0
Published by: Community
Author: newfull5
Last verified: Oct 2026

newfull5/malkuth

User repository on GitHub

View on GitHub

Jev like System One decision models for multilingual

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Malkuth is a small multilingual decision model for classification, with a focus on Korean. It returns probabilities over a fixed set of choices rather than generating text. Two sizes: Malkuth-4B and Malkuth-2B. choice: one of up to 255 options. Each request carries its own criteria, and can contain several independent questions over the same input. The models are post-trained from Kev. They are served with Kev and used through the TypeSafe System One API. The name is motivated by my favorite game.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Ranking

Used inJev for Software & DevTools

Languages: Python

Highlights

Supports choice, noul yes-or-no, and ordered score questions over the same input state.
Post-trained from Kev bases and used through the TypeSafe System One API via Kev serving.
Evaluated on 29 suites and about 129,000 shared items with accuracy and Brier probability scores.
Publishes per-shape latency medians plus per-dataset scores and calibration in the docs.

Watch out

Weights are research use only because the training mix includes non-commercial sources, and benchmark suites keep their own licences. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

open-jev-typed-decision-engine

Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.