How close is it really to the hosted model?

On 308 fresh sealed decisions the README reports 33.1% correct against 36.7% for the hosted model, and the evaluator calls that set unusually difficult. It also states quality on the provider's published real-world workflows has not been measured, so treat your own domain as unknown.

What hardware do I need to run this locally?

About 9 GB of GPU memory in bf16 for the 4B checkpoint and about 19 GB for the 9B. The latency figures are single-H100 and batch 1: p50 13.2 ms on easy and standard items, rising to p50 30 ms and p95 162 ms on hard items with 1-4k-token documents.

Can I point my existing client at it without changes?

Mostly. The server accepts the TypeSafe-style systemone request shape and the same three question types, so existing callers can be redirected. The limits differ: English-only, more than 16 options are split across several passes, and inputs past 16,384 tokens are refused.

Which checkpoint should I start with?

The 4B. The README is candid that this is unsettled: the 4B leads on the hard tier and its calibration, while the 9B wins on the authors' held-out checks but scores worse on public benchmark items. Re-measure both against questions from your own workload.

How do I install and use it?

Install with pip install "jevk5[fast] @ git+https://github.com/allebee/jevk5@v0.3.0", then create JevK5("alibiserikbay/JevK5") and call model.decide(state, question). The 4B needs about 9 GB of GPU memory in bf16.

Apache-2.0, and JevK5 is an independent project not affiliated with TypeSafe AI. The weights are Qwen3.5-4B or Qwen3.5-9B plus a distilled LoRA, merged.

What are its input limits?

It answers with one of 16 letters, so up to 16 options take one pass and more options are handled in several; it refuses inputs over 16,384 tokens. The model is English-only.

CommunityJev Demos & Experiments

Jevk5

Apache-2.0 open-weight stand-in for a hosted decision model: Qwen3.5-4B with a distilled adapter scores noul, choice and score questions through a systemone-shaped server, with no generated tokens.

See highlights

Open github.com

Stars: 147
Language: Python
License: APACHE-2.0
Published by: Community
Author: allebee
Last verified: Oct 2026

allebee/jevk5

User repository on GitHub

View on GitHub

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

Forks: 12
Last pushed: Updated Sep 2026

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

About this project

JevK5 is an independent, Apache-2.0 open-source alternative to TypeSafe's Jev for typed decisions. Give it a state (a ticket, log, policy, or diff) and a yes/no, choice, or score question. It returns a probability for every option in one forward pass, with zero generated tokens. The weights run on your own GPU, and the server accepts the TypeSafe-style /v1/systemone request shape. JevK5 is not affiliated with TypeSafe AI.

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

Tagscommunity jevbest

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

One forward pass returns a probability for every option; the readout is a softmax over answer-letter logits at one temperature.
About 9 GB of GPU memory in bf16 for the 4B checkpoint, and about 19 GB for the 9B.
p50 is 13.2 ms on an H100 for easy and standard items; hard items over 1-4k-token documents run p50 30 ms, p95 162 ms.
JevBench v1.4 places the previous version second of 76 systems and first among open entrants on a sealed set the evaluator calls unusually hard.

Watch out

Independent of the hosted provider and reading option letters rather than its unpublished architecture, so outputs will not match. English-only, capped at 16 options in one pass and 16,384 input tokens, with requests serialized on one GPU and each question evaluated separately. The published benchmark numbers are for the older v0.2 weights; v0.3 has not been run by either public benchmark, and quality on the provider's own published workflows is unmeasured. v0.3 also gives up one standard item against v0.2, judging answers is its weakest family, and the 9B trails the 4B on hard items with worse calibration. 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

celesto

Sandboxed virtual machines for agents: each sandbox is a microVM that starts in about 500 ms, keeps files and state between sessions, and runs locally or in Celesto Cloud behind a CLI and Python SDK.

Jev Demos & Experiments#community#jevbest

Community

CelestoAI/celestoPython

ghosthands

macOS agent loop that moves a real USB HID mouse and keyboard under a vision model, with an experimental decision-model lane picking the next operation and screen target from an element table.

Jev Demos & Experiments#community#jevbest

Community

affirmitv/ghosthandsPython

Jev_Star

StarCraft II research repo combining Jev action selection with optional Astra plans across full-game macro control and SMAC-Hard micromanagement, with videos, replays, and a paper.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

sc2musa/Jev_StarPython

robodiag-harness

Terminal harness that checks ROS 2 robots with deterministic health tests, capability descriptions, SQLite history, and optional Jev evidence routing.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

YueBit/robodiag-harnessPython

jev-ultrafast-mcp

MCP server where one browser_goal call hands a URL, a goal, and code-checked assertions to a server-side loop driven by a decision model over live page elements.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

jiawei686/jev-ultrafast-mcpPython

jev-style

Local Qwen3.5 decision models in 0.8B and 2B sizes with fitted temperatures, systemone-shaped server, playground, guard, MCP tools, and eval harness.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

lawrence3699/jev-stylePython

From the Jev guides

Original write-ups on what this entry is for.

Per-Tick Decisions in Games

Sep 28, 2026

A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.

Jev Demos & Experiments#gaming#agents#control

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai

Jevk5

About this project

Highlights

Questions about this entry

How close is it really to the hosted model?

What hardware do I need to run this locally?

Can I point my existing client at it without changes?

Which checkpoint should I start with?

How do I install and use it?

What is the license?

What are its input limits?

More Jev resources like this

From the Jev guides

From the Jev community