# A hybrid model of Gemma4 with a JEV-like decision head in multi-speaker voice conversation — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/131cc96f067eeeb2

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-09-28 17:42 ingested 2026-09-29 02:43

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1wslcum/a_hybrid_model_of_gemma4_with_a_jevlike_decision/)

## retained text



```
The human brain is neither an LLM nor a JEV. In a crowded market you hear a lot of speech and answer almost none of it. The ongoing question is not “what should I say?” It is “you talking to me?" and "should I say anything at all?”

This live voice demo showcases three hardware tiers—the **Blackwell 4500**, **Jetson Orin NX 16GB**, and **Jetson Orin Nano 8GB**—solving this exact problem. By splitting the workload between a lightweight decision head for turn-taking and a **Gemma 4** pipeline for text generation, the setup delivers highly responsive, low-latency vocal interaction.

EDIT: repo (the latest code yet published) [https://github.com/cortexist/little-gemma](https://github.com/cortexist/little-gemma)
```



document fetched 2026-09-29 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1wslcum/a_hybrid_model_of_gemma4_with_a_jevlike_decision
