shisa.ai
New Qwen4 Max Available! Use with Jouzu — our CLI agentic coding harness, built for Japan. Connect in 60 seconds.Learn More

Shisa DE-2.1

A decision model for Japanese workflows

Ask a question about text, a record, or an image. Get a yes/no answer, a choice, or a score with probabilities your software can use.

Text-based decision

「今回はちょっと難しそうです。」

Is this a polite decline?

Yesyes: 0.91 / no: 0.09
Image + text-based decision
喫茶みどり

Listed name: 喫茶みどり

Does the visible sign match the listing?

Yesyes: 0.94 / no: 0.06

Tokyo latency

~30 ms from central Tokyo

For teams scaling Jev-powered operations in Japan, test Shisa on the same requests and compare median and tail latency at your expected traffic.

Tokyo internal measurements · Hosted APIms
  • Shisa API~30 ms
  • TypeSafe Jev API≥220 ms

Shisa API was measured after optimization. The Jev figure is the median from an earlier hosted API test and was not remeasured.

What it does and where it stops

Yes / no

Check supplied evidence against a criterion. For example, compare a storefront sign with a business listing.

Pick one

Assign a category from up to 256 text options. The SDK supports up to 26 options for image questions.

Score

Rate a case against a scale you define. Your workflow decides when to act and when to send it for review.

It does not generate prose, explain its choices, or verify evidence independently. Image probabilities are not calibrated.

Evaluation

How DE-2.1 compares

DE-2.1 scored higher than Jev 1.13.0 on Shisa’s Japanese nuance and indirect-refusal evaluations.

Shisa evaluations, single read. AUROC measures how well scores distinguish positive from negative examples; it is not accuracy. Methods and source

Japanese nuanceAUROC
  • DE-2.10.974
  • DE-20.965
  • Jev 1.13.00.921
  • OpenAI Decisions0.898
Japanese indirect refusalsAUROC
  • DE-2.10.926
  • DE-20.919
  • Jev 1.13.00.844
  • OpenAI Decisions0.806
Japanese leakage v3AUROC
  • DE-2.10.986
  • DE-20.983
  • Jev 1.13.00.951
  • OpenAI Decisions0.955

Evaluate it on your workflow

Self-host the Apache-2.0 weights with the Python SDK, or talk to Shisa about a hosted evaluation in Japan.