# OpenJev: Run AI decisions locally in your browser

> Compare two ways an AI model makes choices, all on your own device with no server involved.
- Title: OpenJev: Run AI decisions locally in your browser · Openjev
- Summary: Compare two ways an AI model makes choices, all on your own device with no server involved. You can now test how local AI models make decisions directly in…
- Keywords: openjev, local ai, browser-based, decision models, on-device inference, semantic if, minecpm5, qwen3, gpu inference, no backend, ai transparency, technology
- Source: Openjev — https://openjev.com
- Read time: 1 min
- Topics: openjev, local ai, browser-based, decision models, on-device inference, semantic if, minecpm5, qwen3, gpu inference, no backend, ai transparency, technology
## Run AI decisions locally
OpenJev lets you load a decision model directly in your browser and watch it work on your GPU. No server, no waitlist, no backend needed.

> browser only no backend
## Pick your model size
Choose from three options: Qwen3 0.6B for phones, MiniCPM5 2B for desktop, or Qwen3.5 4B for high-memory systems. Larger models download once and stay cached.
## Two decision methods compared
Method one reads choice probabilities directly from the model without decoding. Method two asks the model to write its probabilities as JSON text, token by token.
## Real timing on your GPU
Both methods run on the same loaded model and are timed with actual wall-clock measurements. You see setup time, warmup, execution, and generation speed for your specific hardware.
## Accuracy varies by model
The 0.6B phone model trades accuracy for size. MiniCPM5 reaches 69.3 percent accuracy. The 4B model hits 84.5 percent but needs substantially more memory.
## No calibrated confidence scores
Direct probability scores are normalized only across your supplied options and do not represent true confidence or include every answer the model might prefer.
## Key takeaway

You can now test how local AI models make decisions directly in your browser, comparing fast probability reading against token-by-token generation, with real timing measurements on your own hardware.