Framer AI AgentsDesign and publish professional sites with AI
Promoted
Launched this week
Desert Ant Labs
Small specialized AI models for speech, text, vision
137 followers
Small specialized AI models for speech, text, vision
137 followers
Desert Ant Labs builds small AI models that run on your phone or browser, no internet, no per-use cost. Instead of one big model doing everything, they make small ones, each nailing one task, across speech, text, and vision. Add any model in a few lines of code via one SDK. Free up to 100k monthly active devices.
Desert Ant Labs makes small AI models, each nailing one task, across speech, text, and vision, that run on your phone or browser with no internet and no per-use cost.
Drop any of them into your app with one SDK, in Swift, Kotlin, or JavaScript, just a few lines of code.
Audio Models: Align (word timestamps), Clear (speech enhancement), Clips (clip selection), Ear (spoken language detection), Uhm (filler-word detection), Voz (speech recognition), Who (speaker labeling, beta)
Text Models: Emo (emoji suggestions), Gist (topic tagging), Redact (PII redaction), Title (titles and descriptions), Tongue (language identification), Schemer (structured extraction, beta), Toxic (hate speech triage, beta)
Vision Models: Shapes (shape recognition), Eye (beta), Face (beta), Moderator (content moderation, beta)
What makes it different: Everything runs fully on-device, so there's no cloud bill and no token metering, ever. Free up to 100k monthly active devices per platform, unlimited inference per user after that.
Who it's for: Developers who want speech, text, or vision features without adding cloud costs or latency.
Try it: desertant.com · SDK on GitHub · Models on Hugging Face · Docs
P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends
@rohanrecommends congrats on the lauch its a great producct best of luck .
Macaly
one small model per task, no internet, no per use cost. thats a really usefull bet imo 🐜 whats the most downloaded one so far?
Refocus
The Uhm filler-word model is the one that jumped out at me. I build a voice companion that calls older adults every day, and the hard part has never been transcription, it is endpointing: an 82 year old pauses four or five seconds mid sentence and standard VAD reads that as end of turn and talks over them. A tiny on-device hesitation signal running alongside VAD would be a far better turn-taking cue than silence duration alone.
Two questions: is Uhm exposed as a streaming signal you can read frame by frame, or only after a clip finishes? And do you have any read on how Voz and Clear hold up on older-adult speech, slower rate, dysarthria, hearing aid feedback, a TV going in the background? That population is badly underrepresented in most ASR training data and it is exactly where the cloud models I have tested fall apart.