JevMade

Sign in
← Back to experiments

Benchmarks & research

jev-on-a-laptop

This study benchmarks parallel typed decisions on unmodified 1.5B–8B models running on Apple Silicon.

Source screenshot of jev-on-a-laptop
SOURCE SCREENSHOTFull screenshot ↗

What it does

It includes research notes, full result logs and a Hugging Face Space demo; acceleration may fall back when free VRAM is insufficient.

How you can use it

To try this approach, start by listing your project choices. For example, write down your fraud checks or customer support categories, along with the exact answers each case needs. A developer can then run this study's local code to answer all those questions together using a small model on a computer.

Keep in mind that forcing the model to pick from your list does not guarantee the answer is correct. The test found that the model can be wrong even when it reports very high confidence. You should still have a person review important actions.

Maker-reported (not independently measured by JevMade): 240.9 → 83.8 ms (custom 27 fields). Capture needs free VRAM and may fall back · Also measured: 4-field / 255-choice tariff preset (1.5B: 0.15 s, 5.9x) and support-triage (1.5B: 0.76 s, 4.4x — collision handling costs extra). Full logs and JSON: results/

Primitives
choice
Platform
Python
Added
Project created
GitHub stars
14 (snapshot, not live)

Keep this for later

Sign in to bookmark experiments, guides and videos, and keep notes only you can see.

Continue to sign in

We’ll bring you back to this listing.