Kal

😶

Coming up next 🤗

😶

coming up next 👀

Seeing the words show up while you’re still talking instead of waiting for the recording to finish is such a different experience. Also, it’s open source, so you can grab the weights from Hugging Face, or just try it in the Playground first and see how fast it feels. 😊Playground 😊Hugging Face

Scribe stream is a really fast model. Give it a try on the playground

We used to serve one request at a time. The GPU had room for 24 and we gave it 1. Under traffic only 8% of Amharic requests made it through. Once we let requests share the GPU it went to 99%. Same machine. so that is 24 users using the same model at once @b1n1yamBuilds #inference

Elon Musk bought dot.com before OpenAI launched Dots their always on AI Agent this week and redirected it straight to Grok 😭 @kalBuilds

And this is basically why. It just makes serving models way more practical.

Thank you to everyone who has joined our language evaluator group and started contributing so far. The feedback has already been really useful. In the Tigrinya evaluation alone, people have caught issues around pronunciation, number reading, text quality, and places where the speech doesn’t sound natural to a native speaker. This is why your perspective really matters. We’re still evaluating Tigrinya, Wolaytta, and Sidaamu Afoo, and we’d love to have more native speakers contribute. Thank you again to everyone taking the time to help us improve these models and make them sound more natural and familiar to us. @kalBuids

We’re looking for native speakers of Tigrinya, Sidaamu Afoo, and Wolaytta to help with language evaluations we’re running at Addis AI. If you speak any of these languages natively and would like to participate, join our Addis AI Language Evaluators Telegram group and introduce yourself under your language topic: ✅ https://t.me/+INclR6k6G3FhMmM0 @kalBuilds

People are building some really cool things with the Addis AI API. Sefineh used the chat generation API to build this gradient descent visualization. More dev tools coming your way... @b1n1yamBuilds

One thing I’ll never get tired of is seeing what people decide to build with Addis AI. this one was definitely unexpected. What are you using Addis AI for? @kalBuilds

We’re looking for native Tigrinya speakers to help us with a more detailed Tigrinya evaluation this week. If you’re interested, join our Discord and let us know in the Tigrinya channel: 💰 https://discord.gg/ajuXM4zFA @kalBuilds

My day now a days 😂 @kalBuilds

መልካም አዲስ ዓመት 🌼

@b1n1yamBuilds talks about the early days, our models, open source, and a lot of what happened in between with @software_guy. Worth a listen if you’ve been following the journey.

Speaking at SIMS 2026 this November. I’ll be joining the AI for Good panel to share some of what we’ve learned building Addis AI and talk about what creating meaningful impact with AI looks like in practice. Excited for the conversation. @Kalbuilds

Finally, a clean open Amharic speech dataset 🇪🇹🎙️ We ran it through 7 ASR systems to stress-test the labels. it held up clean, even the best models track the references tightly (CER 8–14). the ground truth is legit. 😊Read the full blog @b1n1yamBuilds

Been waiting to share this one. The listening test was fully blind, so nobody knew which model they were hearing or which one was ours. Same text, no names, just Amharic speakers rating what they heard. Makes the results a lot more interesting when nobody knows who they’re supposed to be rooting for. Full results: 😊addisassistant.com/benchmarks/addis-voice-2-amharic Dataset: 😊 huggingface.co/datasets/addisai/amharic-tts-benchmark @kalBuilds