Is local accuracy competitive with hosted Jev?

Close but behind: the default model scores lower than hosted Jev on the public JevBench set, roughly on par with SemIf, with the gap concentrated on the hard tier. Full results ship in the docs.

Will my existing TypeSafe client code break?

No changes needed. Point the SDK at the local base URL with any key value and the same state plus noul, choice, and score shapes work, including per-request model picks.

What happens when my state exceeds the local window?

It keeps the start and end and marks the response with a truncation header instead of refusing, while the Jev-shaped body stays intact for your code to handle.

Can I run this fully offline after setup?

Yes after the first model download; the small model fetches first and the large one only on request, then everything stays on the machine with no outbound calls.

How do I install and run it?

Clone the repo, create a virtual environment with uv (Python 3.10 or newer), install with uv pip install -e ., then run .venv/bin/local-jev serve, or add --ui for the browser portal. Both listen on http://127.0.0.1:8765.

What models can I choose and how big are they?

The smallest is nli-deberta-large at 0.87 GB, an entailment model; the default and most accurate is llm-qwen3.5-4b at 9.32 GB on disk and about 8.5 GB of memory. The first start downloads only the smallest and never starts a large download by itself.

What license is local-jev under?

MIT, and the author describes it as a research project rather than a finished product. On JevBench's 231 public items the default model answers 80.5% correctly, against 86.6% for the hosted Jev API.

CommunityJev Practices & Patterns

Local Jev

Local offline System One server speaking the hosted /v1/systemone wire format, so existing SDK code runs unchanged against small on-machine models.

See highlights

Open github.com

Stars: 17
Language: Python
License: MIT
Published by: Community
Author: amithgc
Last verified: Oct 2026

amithgc/local-jev

User repository on GitHub

View on GitHub

A local, offline System One server compatible with TypeSafe's Jev API. It answers typed yes/no, category and score questions with small open models.

Forks: 2
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A local, offline System One server. Software that needs a decision rather than prose (which queue, how severe, is this spam) sends a piece of text and some typed questions, and gets back a choice, a score or a yes/no, each with a probability for every possible answer. local-jev speaks exactly the wire format of TypeSafe's hosted Jev API, so the official SDK and existing client code work unchanged, but the models run on your own machine: small open models, a Qwen instruction-tuned model by default and an…

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

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Ranking

Used inJev for Software & DevTools

Languages: Python

Highlights

Defaults to a Qwen instruction model with a lightweight entailment option onboard.
Serves POST /v1/systemone plus models, health, and docs on 127.0.0.1:8765.
Accepts per-request model choice with jev-latest aliases mapping to the default.
Shortens overlong states with a truncation header instead of refusing.

Watch out

A research project, not a finished product; CPU runs slowly so a CUDA or Apple GPU is advised, and portal routes stay unprotected so keep it on loopback. 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

jev-rag-benchmark

Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.

Jev Practices & Patterns#community#jevbest#catalog-import

Community

erendikmenn/jev-rag-benchmarkPython

jevshield

Framework-agnostic guard layer that puts Jev in front of an agent's tool calls: a Choice risk tier, a Noul destructiveness check and a Score blast radius decide allow, deny or ask.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

lgy1027/jevshieldPython

jev-mcp-server

Python MCP server exposing Jev's three official question types plus compare, verify and batch classify, with an installer subcommand that writes the client config and a measured cost per call.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

wangkuangkuang/jev-mcp-serverPython

qwen-rlcd

An inference-only prototype of a Jev-style decision model on a sub-billion-parameter hybrid backbone, returning typed probability distributions in one forward pass, with its architecture treated as an explicit guess.

Jev Practices & Patterns#community#jevbest#catalog-import

Community

shamazharikh/qwen-rlcdPython

typesafe-vs-deepseek

A Vercel-hosted comparison of TypeSafe (Jev) against DeepSeek-flash on speed, tokens, cost and accuracy, backed by per-task Python pipelines, fixture generators and a report script. Eighteen commits, nine issues still open.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

markfive-proto/typesafe-vs-deepseekPython

what-is-jev

Independent research repository on Jev: sourced findings that separate verified claims from vendor marketing repeated by others, a rubric-scored catalog of public Jev repositories, and recurring design patterns with evidence.

Jev Practices & Patterns#community#awesomejev#catalog-import

Community

g0runmezadam/what-is-jevPython

From the Jev guides

Original write-ups on what this entry is for.

Designing the State: What to Include and What to Drop

Oct 12, 2026

Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.

Jev Practices & Patterns#state#design#primitives

Read article

Noul, End to End: One Probability and a Dozen Thresholds

Oct 12, 2026

A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.

Jev Practices & Patterns#noul#primitives#confidence

Read article

Scoring Without Measuring: Bands, Rubrics, and Composite Triage

Oct 12, 2026

A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.

Jev Practices & Patterns#score#primitives#scoring

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai