What is JevAlt and which models does it ship?

JevAlt is a free and open-source project with three 4B open-weights decision models: Deem-4B for English, Karar-4B for Turkish and Wähler-4B for German. Given a situation, a question and answer options, each model returns a probability per option instead of generating text.

How does JevAlt relate to Jev and TypeSafe?

Its server implements the Jev API, POST /v1/systemone, so existing Jev clients can use it, and the API follows TypeSafe's public System One specification. The README states JevAlt is independent of TypeSafe AI; Jev is a TypeSafe AI model.

How do I run JevAlt locally?

Install with pip install "jevalt[serve,gguf] @ git+https://github.com/mertkayacs/jevalt" and run jevalt serve; it downloads Deem-4B and starts a server at http://127.0.0.1:8000. The Q4_K_M GGUF builds run on a CPU in about 3 GB of RAM, with a GPU optional.

What accuracy do the JevAlt models report?

On held-out decisions the README reports 94.7% for Deem-4B English, 96.8% for Karar-4B Turkish and 92.0% for Wähler-4B German, each compared with Kev-4B. It notes the tests come from JevAlt's own data pipeline and favour it.

Does JevAlt need an API key or payment to run?

No: the models are self-hosted and Apache-2.0, released from LoRA training on one A100 80GB. The README reports about 2.7 A100 hours for released training runs and 32.2 USD total project cost including labeling.

What limits and weaknesses does JevAlt document?

Confidence is calibrated on held-out project data and should be refit on your own data before setting thresholds; long noisy text and date calculations are weak spots; and the models still follow some hidden instructions planted in input text, so test that behaviour before triggering actions.

Who built JevAlt and what license applies?

Mert Kaya of Eschatia Labs publishes JevAlt under the Apache-2.0 license for both code and weights, starting from Intern-Decision-4B. The repository was created 2026-10-01 and pushed 2026-10-16, with Hugging Face weights for all three models.

Jev directory

CommunityJev Models & Reimplementations

JevAlt

Three open-weights 4B decision models for English (Deem-4B), Turkish (Karar-4B) and German (Wähler-4B) that choose between options and return a probability for each. The server implements the Jev API, POST /v1/systemone, so existing Jev clients work by changing the endpoint.

See highlights Open github.com

Stars: 5
Language: Python
License: APACHE-2.0
Reading time: 2 min
Published by: Community
Author: mertkayacs
Last verified: Oct 2026

mertkayacs/jevalt

User repository on GitHub

View on GitHub

Three 4B decision models for English, Turkish and German, with per-option probabilities, GGUF builds and the Jev API.

Forks: 1
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

Tagscommunity github python gguf multilingual

Works forJev for Classification Jev for Routing Jev for Decision gating

Used inJev for Software & DevTools Jev for Customer Support

Languages: Python
Platforms: API CLI
Capabilities: Open sourceSelf-hostedRuns locally

Highlights

Three 4B models (English Deem, Turkish Karar, German Wähler) LoRA-trained from Intern-Decision-4B.
The server implements the Jev API POST /v1/systemone; Q4_K_M GGUF builds run on a CPU in about 3 GB of RAM.
Held-out accuracy reported as 94.7% English, 96.8% Turkish and 92.0% German, each against Kev-4B.
Live tests: Deem answered 122 of 130 English requests correctly across five cases on 2026-10-04.
Apache-2.0 code and weights; project cost reported at 32.2 USD and about 2.7 A100 hours for released runs.

Quickstart

bash

pip install "jevalt[serve,gguf] @ git+https://github.com/mertkayacs/jevalt" && jevalt serve

Watch out

README reviewed 2026-10-17; the README says accuracy is measured on the project's own data pipeline and favours JevAlt, calibration should be refit on your own data, and the models still follow some hidden instructions planted in input text.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

X: Both Jev 1.13 and 1.14 score 29 of 30 on 30 Turkish buyer and seller messages

From the Jev guides

Original write-ups on what this entry is for.

Beyond text and boxes: multimodal decisions you can run yourself

Oct 16, 2026

Jev-Omni's merged classifier scores text, image, audio, and video options; VideoJev and Laya Vision show what open multimodal decision models accept, what they return, and which numbers are the authors' own.

Jev Models & Reimplementations#multimodal#open-models#vision

Read article

Jev 1.14: Images, Box, and the Danger Zone

Oct 15, 2026

Jev Day shipped a routine model bump and a deliberate widening of the API: images on both models, a box question type on 1.14 only, and 262k context, all behind danger-zone headers.

Jev Models & Reimplementations#release#vision#danger-zone

Read article

The Open Decision-Model Roundup

Oct 14, 2026

Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.

Jev Models & Reimplementations#open-models#training#local

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai

JevAlt

Highlights

Quickstart

Questions about this entry

What is JevAlt and which models does it ship?

How does JevAlt relate to Jev and TypeSafe?

How do I run JevAlt locally?

What accuracy do the JevAlt models report?

Does JevAlt need an API key or payment to run?

What limits and weaknesses does JevAlt document?

Who built JevAlt and what license applies?

Reactions & coverage

Related terms

More Jev resources like this

From the Jev guides

From the Jev AI model community