Skip to content
OPEN MODELS. YOUR MACHINE.

See. Decide.
Stay local.

Turn text and images into typed probabilities. A Jev-compatible API with one output token per question, running on your Mac.

$uv run openjev serveApple Silicon + Metal
INPUT / 输入
AaText & context
Images & frames
SystemOne1 TOKEN
{
  "type": "choice",
  "probabilities": { … }
}
Typed. Composable. Local.
DECISIONS AS A PRIMITIVE
1 token

per question · no generated reasoning

3 primitives

Noul · Choice · Score

255 options

complete label distributions

100% local

inference stays on your machine

01 / FROM INPUT TO ACTION

A small API.
A useful kind of intelligence.

Route a message. Read a screen. Choose the next step. The response is already structured for your code.

RECORDED LOCAL RUN
STATEQwen3.6-35B-A3B

Loading recorded result…

RESPONSE1 OUTPUT TOKEN
choice
measured HTTP latencyinput tokens
Replay of an actual local API response. Run the playground on localhost to try your own inputs. Scores are conditional on the supplied options. Try it →
02 / PROGRAM WITH PROBABILITIES

Three ways to make a decision.

01

Noul → float

Is it true? A yes/no judgment expressed as a probability you can threshold.

if refund.noul > 0.9:
  review_payment()
02

Choice → enum

Which one? Select an option and inspect the complete distribution.

route = answer.choice
dispatch[route](ticket)
03

Score → float

How much? Get the expected position on your ordered rubric.

priority = urgency.score
queue.add(item, priority)
03 / MEASURED, ON A MAC

Show the work.

Nine tasks. Our API. Reproducible inputs, per-question receipts, and a clear account of what was measured.

Measured OpenJev Multimodal benchmark scores with sample counts; see the accessible data table on the benchmark page.
Zero-shot · One-token readout · 20 samples per task · Seed 42Methodology & raw results ↗
BUILT FOR YOUR NEXT IDEA

Your state.
Your model. Your decision.

Start with a lightweight 4B model. Scale to Qwen3.6 on a larger Mac. Keep the same API.

Open models. Local inference. Measured claims.