unusual_typo

Here are the benchmark results. You can check more details in the repo. openai/privacy-filter on Apple M1 Max

   dtype              1k total    1k tok/s       8k total    8k tok/s
  ━━━━━━━━━━━━━━━━  ━━━━━━━━━━━  ━━━━━━━━━━  ━━━━━━━━━━━━━  ━━━━━━━━━━
   fp32              620.52 ms       1,664    4,893.86 ms       1,689
  ────────────────  ───────────  ──────────  ─────────────  ──────────
   fp16              654.56 ms       1,578    5,430.17 ms       1,521
  ────────────────  ───────────  ──────────  ─────────────  ──────────
   q4                582.13 ms       1,776    4,635.39 ms       1,784
  ────────────────  ───────────  ──────────  ─────────────  ──────────
   q4f16             648.10 ms       1,594    5,261.56 ms       1,570
  ────────────────  ───────────  ──────────  ─────────────  ──────────
   quantized int8    573.94 ms       1,801    4,594.95 ms       1,800

I just ran the benchmark on my macbook. 582 ms for 1k tokens and 4.64 s for 8k

It use openai/privacy-filter which is smaller than 1GB in size. I haven't checked usage during inference. It rans at 1k toks/sec on my macbook. I will update the repo with benchmark results. Thanks for the comment

I wanted to implement the feature initially. i realized that it requires modification of coding agents (eg codex, claude code, opencode etc). hook or skills pass PII data into server eventually so i decided to share the standalone app first. Feel free to submit a PR!

I shipped an enhanced fork of Parlor (by Fikri Karim https://news.ycombinator.com/item?id=47652007) that reads various visual inputs and uses Supergemma 4 E4B + Supertonic TTS to run a fully local, multimodal, and multilingual AI assistant. It runs entirely on your machine.

What it does:

1. Talk to your screen: It reads and understands your webcam, screen sharing, PDFs, and video at once.

2. Native Multilingual: It can speak five languages: English, Korean, Spanish, Portuguese, and French.