things i'm proud of
in no particular order.
kea-b
A 0.6B-parameter speech recognizer whose encoder stores every large matrix at one bit per weight, +1 or −1. 105 MB, 23x smaller than its parent. It trades 0.6 points of word error rate for 33% lower latency and about 37% less energy per clip on an RTX 5090. Press the button. Your audio never leaves this tab.
Runs inside this browser tab.
The model is 105 MB. Best on Wi-Fi and a laptop or desktop.
Downloading weights
0 / 105 MB
huggingface.co/rajb3/kea-b
Record
0:00 / 0:30
30-second limit
Transcript
This browser doesn't expose WebGPU, so Kea-B can't run here. Try a recent Chrome, Edge or Safari on a laptop or desktop.
Kea-B transcripts of the sample recordings
- 01Download105 MB · Hugging Face
- 02VerifySHA-256
- 03CompileWebGPU kernels
Audio stays in this tab
rajb3/kea-b · CC-BY-4.0
- localflowthe dictation app i use every day. kea-t on a mac's neural engine, through Core ML, with iOS and Linux versions. nothing leaves the machine.
- sage.nviman llm inside neovim, with the editor's diagnostics as context and a retrieval pipeline over the repository.
- kea-tthe ternary model kea-b grew out of: 98% of its weights in {−1, 0, +1}, 181 MB, 6.84% mean word error rate against the original's 6.45%.