Pinned Tweet
Today we're releasing Underdog Saluki 27B
Based on Qwen 3.8 27B, Saluki is almost 7x smaller than its full-precision counterpart while it keeps 96% of its benchmark performance and beats Qwen at tool calling
It's the best model <8GB that runs in Underdog
Saluki is available today under Apache 2.0. Try today in Underdog underdog.ai/saluki
Underdog AI retweeted
For absolutely no reason, save the fact that I can burn tokens locally, and @loyal allows me to do it, here's two pictures from the default prompt of a local AI model called Poodle 😂
Underdog AI retweeted
■ わずか40.5MBの日本語音声AIが、大型モデルより読み間違いが少ない!?
超軽量TTS「Underdog-Woof-Japanese」が公開された。
GPU不要、API不要。ノートPCのCPUだけで日本語音声を生成できる。
■ 大型モデルとの比較
難しい日本語281文の読み間違い率
・Underdog Woof(40.5MB):1.33%
・Fish Audio Drama 3(41.5GB):2.39%
・Qwen3-TTS 1.7B:8.88%
・Kokoro:11.27%
約1,000倍のサイズがあるDrama 3よりも、読み間違いが少ないという結果。
■ なぜこんなに小さい?
日本語の読み方をAIに全部判断させず、Open JTalkで読み仮名やアクセントを解析。
40.5MBのモデルは音声生成を担当する。
つまり、既存技術とAIを分業させている。
■ 動作性能
・生成速度:実時間の14.5倍
・最大メモリ:437MB
・NumPyだけで動作
・音声は2種類(Misa / Mikasa)
ローカルLLMの音声出力にも使えそう。
数値は開発者による測定で、第三者検証はまだない。
商用利用は別途許諾が必要。
Today we’re releasing Underdog Woof in Japanese. Woof が日本語を話せるようになりました。
Two voices in a 40 MB model, Misa & Mikasa, that runs 14.5x faster than real time
Woof JP generates speech fully in your browser, even without wifi. Try it in our game: underdog.ai/escapejp
It’s free and open weight on huggingface. Try Woof today in Underdog, your Private Personal AI: underdog.ai/woofjp
Today we're releasing Underdog Woof, the #1 frontier TTS model
It's only 31.7 MB and beats Text-to-Speech models 1000x larger on Word Error Rate
It's small enough to run fast in your browser, phone, watch... no data centers
It’s free and open weight with Underdog CC BY-NC 4.0. Try Woof today in Underdog, your Private Personal AI: underdog.ai/woof
Play our game, where Woof JP run’s 100% in your browser in WebGPU and works without wifi
Try to defeat Darkcloud and free Riverbark from Orwellian Big AI: underdog.ai/escapejp
2
1,254
Underdog AI retweeted
local models are secure models
Today we’re releasing Underdog Woof in Japanese. Woof が日本語を話せるようになりました。
Two voices in a 40 MB model, Misa & Mikasa, that runs 14.5x faster than real time
Woof JP generates speech fully in your browser, even without wifi. Try it in our game: underdog.ai/escapejp
It’s free and open weight on huggingface. Try Woof today in Underdog, your Private Personal AI: underdog.ai/woofjp
Underdog AI retweeted
ここかなり偉い
Today we’re releasing Underdog Woof in Japanese. Woof が日本語を話せるようになりました。
Two voices in a 40 MB model, Misa & Mikasa, that runs 14.5x faster than real time
Woof JP generates speech fully in your browser, even without wifi. Try it in our game: underdog.ai/escapejp
It’s free and open weight on huggingface. Try Woof today in Underdog, your Private Personal AI: underdog.ai/woofjp
Underdog AI retweeted
Just switched to @loyal and wow I'm impressed.
Everything is private, on-device consumer AI that’s free.
No longer need to pay for Superhuman or Granola. It's just a native app within Underdog. Feels the exact same.
Trying out the Todos and Speak (wispr dictation) features next 👀
Underdog AI retweeted
Today we're releasing Underdog Woof, the #1 frontier TTS model
It's only 31.7 MB and beats Text-to-Speech models 1000x larger on Word Error Rate
It's small enough to run fast in your browser, phone, watch... no data centers
It’s free and open weight with Underdog CC BY-NC 4.0. Try Woof today in Underdog, your Private Personal AI: underdog.ai/woof
Underdog is your private personal AI that works on devices you already own
Underdog AI retweeted
うぉー!熱い🔥
小さい容量で高性能、文字から音声に変換するモデル(TTS)が出た!!
Underdogの「Woof」ってやつ
なんと31.7MBしかないのに、1000倍大きいモデルを超える精度らしい
ブラウザでそのまま動くし、Wi-Fi切っても使える。ブラウザゲームとかAIチャット系で使えそうだな
※対応言語は英語のみらしい。日本語は難しいか
オープンウェイト(非商用)で公開されてる
デモゲームのクオリティ高過ぎて最後までやってしもた…
underdog.ai/woof
underdog.ai/escape
Today we're releasing Underdog Woof, the #1 frontier TTS model
It's only 31.7 MB and beats Text-to-Speech models 1000x larger on Word Error Rate
It's small enough to run fast in your browser, phone, watch... no data centers
It’s free and open weight with Underdog CC BY-NC 4.0. Try Woof today in Underdog, your Private Personal AI: underdog.ai/woof
Underdog AI retweeted
I like using @loyal to check status of my email, meetings, calendars, etc. everything stays local and this gives me a sense of control and freedom.
P.S.: yes I have some invites available as you can see in the screenshot.
Underdog AI retweeted
My M1 MacBook Air sat untouched since college. Underdog just brought it back to life — Reorganizes my email and sits in my work meetings transcribing and summarizing everything. All on-device, on a tiny 4B model. This 2020 laptop is useful again.
Replying to @loyal
Retweet to get your friends & family aware, protected & an invite to use Underdog
Safety is the priority. Free, safe, & smart AI for everyone underdog.ai/?v=42
Underdog AI retweeted
sent dl.new to a friend and he said he wasn't comfortable using it since he's been "pwned 3 times the last one month with railway, trigger and one more".
i told him my repo is open source and he said "php got hacked idk anymore man".
made me think how i'm supposed to think about trust/safety as a founder given the AI hacks. I imagine there will soon be companies that use AI for daily security testing and certification.
already seeing a glimpse of that future on websites like @loyal 's but i think this will be a huge market within months.
Underdog AI retweeted
Underdog team is releasing at hypersonic speed! 🚀
Today we're releasing Underdog Woof, the #1 frontier TTS model
It's only 31.7 MB and beats Text-to-Speech models 1000x larger on Word Error Rate
It's small enough to run fast in your browser, phone, watch... no data centers
It’s free and open weight with Underdog CC BY-NC 4.0. Try Woof today in Underdog, your Private Personal AI: underdog.ai/woof
Underdog AI retweeted
ワイの好きなUnderdogAIちゃんのTTSモデルがリリースされたようです!
英語オンリーですが31.7MBっていうとんでもないミニサイズ
あとWebサイトのデザインがクソ可愛くていい
Today we're releasing Underdog Woof, the #1 frontier TTS model
It's only 31.7 MB and beats Text-to-Speech models 1000x larger on Word Error Rate
It's small enough to run fast in your browser, phone, watch... no data centers
It’s free and open weight with Underdog CC BY-NC 4.0. Try Woof today in Underdog, your Private Personal AI: underdog.ai/woof
Woof is 31.7 MB, so small that the whole model runs in your browser in WebGPU
You can generate voices even after turning off Wifi
Try woof here: underdog.ai/escape
Today we're releasing Underdog Woof, the #1 frontier TTS model
It's only 31.7 MB and beats Text-to-Speech models 1000x larger on Word Error Rate
It's small enough to run fast in your browser, phone, watch... no data centers
It’s free and open weight with Underdog CC BY-NC 4.0. Try Woof today in Underdog, your Private Personal AI: underdog.ai/woof





























