This AI Company Is Making The Most Underrated Models (Desert Ant Labs)

Author: Better Stack · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: This AI Company Is Making The Most Underrated Models (Desert Ant Labs) · Channel: Better Stack · 15:35 · Watch on YouTube

Topics: AI infrastructure and hardware, AI models, Local and open-source models, Multimodal models

Summary

Desert Ant Labs is a European AI lab focused on small, specialized models that run offline on phones and laptops, handling fast background tasks while larger cloud models do the reasoning. The video tests 12 models covering speech, text, and vision, with mixed results: some are useful and efficient, while others struggle with accuracy or specific inputs. The presenter highlights the models’ small footprints, open availability, and potential for app development. He sees on-device micro-models as a promising direction for AI.

Key points

Desert Ant Labs builds small on-device models intended to work offline without servers, API keys, or token costs.
Its approach is inspired by the cerebellum: specialized models handle quick, routine tasks while larger models handle reasoning.
The presenter reports that Voz transcribed a seven-minute recording in about two seconds.
Clear reduced noise in a difficult kitchen recording, though the presenter felt the result was only partly successful.
Clips automatically transcribes long videos and ranks short clips, but the presenter found its rankings difficult to adjust and not always ideal.
Redact recognizes some personal information but failed to detect a password and a Canadian SIN number in the test.
Emo and Tongue stood out to the presenter for their small sizes and fast responses.
Other tests found limitations: gist misclassified topics, Shapes struggled with an ellipse, and Ear misidentified spoken Latvian as Turkish.
The lab’s models are open on Hugging Face, its SDK is open on GitHub, and use is free up to 100,000 monthly active devices per platform.
The presenter views efficient, task-specific on-device models as a promising direction for future apps.

Chapters

Start	Chapter	Summary
00:00	Desert Ant Labs and its on-device AI philosophy	The presenter introduces the lab’s specialized offline models and explains how they are meant to handle quick background tasks.
01:32	Speech models: Voz and Clear	The presenter tests Voz transcription speed and Clear’s ability to remove noise from a kitchen recording.
04:09	Clips, Redact, and filler-word removal	The video evaluates automatic clip selection, personal-data redaction, and the model for removing filler words.
07:47	Emoji suggestions, topic tags, and word alignment	The presenter tries Emo and gist, then describes Align’s word-timestamp function and its testing limitations.
10:25	Shapes, Title, and language identification	The video tests shape recognition and language identification, including Tongue and Ear, with mixed results.
14:03	Beta models and the future of on-device AI	The presenter lists models still in beta and reflects on the potential of efficient, task-specific AI.

Quotes

they’re building tiny models that run entirely on your phone or your laptop offline — 00:00

it did actually transcribe it in 2 seconds. — 02:04

it’s only 5 MB on device — 08:19

I wouldn’t say this is production ready, to be honest. — 09:55

I think it’s so cool that there are companies like this one that are focusing on models that are super good at one specific task — 14:34

Concepts and entities

Desert Ant Labs, Voz, Clear, Clips, Clipper, Redact, Um, Emo, gist, Align, Shapes, Title, Tongue, Ear, Who, Shammer, Toxic, Moderator, I and Face, Hugging Face, GitHub, Claude Sonnet

Full transcript (original language)

00:00 This might genuinely be the coolest AI company out there. It’s called Desert Ant Labs and they are a European AI lab and they’ve set out on a very specific mission. Instead of chasing building big models, they’re building tiny models that run entirely on your phone or your laptop offline with no server and no API keys required and with no token cost. So in this video, we’ll take a look at 12 of their little AI models, see how they work and test them out one by one. It’s

00:30 going to be a lot of fun, so let’s dive into it. So Desert Ant Labs whole philosophy is built around a certain body part, the cerebellum. It’s about 10% of your brain by size, but it holds 80% of your brain’s neurons. And that’s the part of your brain that handles the fast thinking and reflexes like balance, timing, coordination, so the rest of your brain is free to think. And Desert Ant Labs wants to apply the same layer

01:02 for software. So what they’re doing is building small specialized models that handle the constant background work on device while the big cloud models handle the actual reasoning. And so far they’ve shipped 12 of these for various tasks like converting audio, text and vision and they have a handful of beta models coming out soon as well. And look at their marketing pages. Look how gorgeous each of them are. I’m just so blown away by how stylish and tasteful their model presentations are. And everything

01:32 they’ve shipped is open on Hugging Face, their SDK is open on GitHub and it’s free to use up to 100,000 monthly active devices per platform. So I’m going to try out every single one of these 12 models and see how they perform. So let’s go through the lab. First up, we have Voz. That’s their special speech recognition model. The claim on their site is that they can transcribe 10 minutes of audio in about 2 seconds running entirely on an iPhone.

02:04 So, let’s see if that claim actually holds up. So, for this purpose, I vibe coded a little app using Opus 5 that uses this Voz model under the hood. And here, I’m going to select a file and I recently visited Fan Expo where I saw Mr. T in person and I recorded some clips of him talking. So, I’m going to use one of these clips which is a 7-minute long video and see how well it transcribes it. So, and as you can see by the numbers here, it did actually transcribe it in 2 seconds. And here’s a

02:35 little playback of what that transcription actually looks like. >> A black kid in the White House. You know, what is the joke or the main thing? You know, tell people I wasn’t an uninvited guest. I didn’t jump over the wall. I wasn’t out there protesting. >> Next, we have a model called Clear. And this model promises speech enhancement that gives you warm, closed-mic’d sound of a podcast studio. So, let’s try to

03:06 record some awful, noisy audio and see how well it processes it. I vibe coded a little iPhone app so I can test out the model directly on my iPhone device and I recorded a little sound bite in my kitchen and I turned the fan up so you can hear how that sounds in the original recording right over here. >> Hello, I’m coming to you live from a very, very loud kitchen in my apartment. Let’s see how this all sounds after

03:38 post-processing. >> Indeed, let’s see how that sounds. So, with Clear applied, it sounds like this. Hello, I’m coming to you live from a very, very loud kitchen in my apartment. Let’s see how this all sounds after post-processing. So, I mean, it’s okay. I feel like the noise removal was 70% there, but to be honest, I gave this model a really hard task.

04:09 So, I would assume for a slightly less noisy audio, you would have a better output, but I just wanted to test it out on the most extreme example. Okay, next up we have Clips. And this is a model that turns long recordings into cherry-picked shorts with highlights from the video’s transcript. And this one has an insane benchmark because they claim Clips was able to clip 100,000 30-minute videos using 470 times less

04:39 energy for the same job as Claude Sonnet. And they also ship this model with a desktop app called Clipper. So, for this demo, I’m going to use Clipper and see how it works. So, the app is quite intuitive. You just drag in a video and then it processes it, transcribes it, and cuts it into short clips. And these clips are ranked based on what the model decides are the best moments of the whole video. And since I know the content of this video, I have to say their ranking is not the best

05:09 because in terms of virality, there were better moments for sure that should have been ranked higher. And it’s a shame that there’s no way to tweak the way the model ranks the clips. But anyway, you also get a little title for each clip and a summary paragraph. And this comes from their little title model, which we will check in a moment. Next up, we have Redact, which is a model that strips personal data as you type. So, sensitive information never

05:42 reaches servers or logs. And it supports 27 languages. And this one has a ready-to-use demo on their marketing page, so I’m going to use that for this test. All right, as you can see, we already have some kind of an example here. It’s highlighted, and we can see how it looks redacted. So, let’s start typing something in. My name is Andres. Let’s try another one. My password is hello, something something. And look at that, it doesn’t recognize passwords.

06:12 So, that’s a bummer. I think that should be included in the model, but nonetheless, what if I say, “My SIN number is”? And then, let’s write something in. So, as you can see, it doesn’t even recognize a Canadian SIN number. So, that’s a bit disappointing. Let’s try my native Latvian language. And as you can see, it also understands my native language, which is pretty cool. All right, let’s move on.

06:42 Moving on to the next one, which is um, and this model claims to remove all filler words within 20 milliseconds, and it can process an hour of audio in just 12 seconds. So, once again, I have installed this in a Vibe Coded app, just so I can test it out on my iPhone. Um, so, um, I am I am working on a video today, um, and it’s I I guess it’s going pretty well. Um, I am now recording an audio for a,

07:16 hm, a certain portion of the video. So, it found five fillers, and it did this in under a second. So, I am I am working on a video today, and it’s I guess it’s going pretty well. I am now recording an audio for a, hm, a certain portion of the video. Let’s see how that looks. So, it did not clip my, hm, portion, but it did remove most of the ums. So, I would say it’s a nice model.

07:47 I wouldn’t say it’s perfect, but it does its job. And I’m going to be honest, I wasn’t a fan of the cut as well, because it didn’t slice the audio neatly. Nonetheless, it’s a cool little model as well. Next one is called Emo, and this is probably one of my favorite ones. It basically suggests an emoji that fits the text you’ve just written, and they claim it can do it faster than you type. So, let’s check how this works. So, we have a bunch of examples here. Book a

08:19 flight to Tokyo, and it shows a plane. Let’s do a country, Canada. It gives me a maple leaf, so pretty close. I’m surprised it doesn’t give you a flag, but okay. I’m running late. Running late, so we have a running man, clock, another clock. I guess for late, it suggests a moon emoji. So, I think this is a pretty fun little model, and for most of the part, it’s quite accurate. And it’s only 5 MB on device, and it gives you a suggestion in under 2 ms.

08:51 So, that is a pretty low memory footprint to have on your apps. Next up, we have gist, and this is a model that basically predicts topics and tags based on the titles of your post. And it claims to be able to do this in 101 languages, and they also have an on-page demo. So, as you can see, we have an example here of an English title, and it recognizes that full body workout means health and fitness with

09:23 99% accuracy. So, here’s a a news article I just pasted from a news site, and it talks about dogs, whether they’re sad or scared. Study suggests and it suggests arts and culture, which is not actually accurate because this is a scientific study, but okay. Here’s another one that talks about space travel, and this gives me careers and productivity. Not the most precise result I was hoping for. I think this model needs some improvement because I

09:55 think this model mainly focuses on keywords and not on context, so I wouldn’t say this is production ready, to be honest. Moving on, the next one on our list is Align, and this is a small model that generates word timestamps on device from any transcriber, and they claim that it is more accurate than Apple’s own transcriber because it is able to detect the actual beginning and the end of a word more precisely. And this one is a

10:25 bit tricky to test because in order for me to run a nice visual demo, I would need to generate two separate transcripts of an audio and video file, and then try to see if it matches precisely with the millisecond precision. So, for this one, I’ll just trust their claim here that it’s pretty performant, and I’ll let you folks try it out for ourselves. But let’s move on to the next model. So, next up we have Shapes, and this is an on device shape recognition model

10:56 that turns your hand-drawn strokes into precise vector graphics. And I got to say it’s it’s nice. I mean, it does recognize a line, a rectangle, a triangle, and a star, but I couldn’t get it to recognize an ellipse properly if I draw it with two separate lines, only if it’s like a circle line. So, this model is cute and all, but I think it still needs some improvement. Let’s move on to the next one, which is

11:26 called Title. And this one basically generates a short factual title and a short description based on a text block or in this case with Clipper as we saw before, it can suggest titles for your short clips based on the text content. We already saw it how it works in Clipper and it’s pretty cool. So, I’ll just leave it at that and let’s move on to the next model. Next one is called Tongue and this model is able to identify the written language

11:57 from just a few words and it does it in tens of microseconds and it supports up to 84 languages. And they also have an on-page demo here, so let’s test it out. Okay, we have some examples here. Je voudrais un café. It’s French. Mitsubishi, it seems to think this is Polish. Not quite, but solid effort. Let’s see if it recognizes Latvian. It does recognize Latvian in this case, which I’m super happy about. I have no complaints.

12:27 Tongue works pretty good. And this model is only 2 megabytes in size. So, I think the memory footprint is so small that this is probably worth putting in your application if you have a production need for it. And last but not least, we have Ear, which is a similar concept as the Tongue model, but this one can identify 99 languages from a short stretch of audio. So, for this one, I’m going to be using my Vibe Coded app as well and let’s

12:59 start talking and see if it can recognize this is indeed English. And yes, indeed, it is English with 100% accuracy and it identified it in 10 seconds. So, I would say it is a bit slow, but it does the job. Let’s try different languages. Šodien āra ir ļoti skaists laiks. And let’s see how well it does with my native Latvian language. Okay, this is not good. The model took 5 seconds to process the audio and with

13:32 29% confidence, it thinks it’s Turkish when in reality it is not Turkish. It’s I don’t think Latvian is even similar to Turkish. So, I’m quite disappointed in that to be honest. But, okay, let’s move on. Let’s try French. Bonjour, ça va ? Je voudrais un croissant. And again, 5 seconds for processing and in this case it did identify that it’s French with 96% accuracy. So, at least it works for some languages.

14:03 So, there you have it. We’ve gone through 12 of their recently published models. And Desert Ant Labs also have five models still in beta. Who is for speaker identification, Shammer is for pulling structured JSON out of raw text, Toxic is for catching hate speech before posting, Moderator for flagging nudity before upload, and I and Face for different vision tasks. These aren’t fully released yet, but seeing how fast this little AI lab is pumping out models, I’m not surprised if they’re

14:34 going to be available very, very soon. So, there you have it, folks. That is a little introduction into Desert Ant Labs. I really think they’re going for the right approach here because in the world with so many frontier AI models that require a ton of disk space and a ton of RAM to run, I think it’s so cool that there are companies like this one that are focusing on models that are super good at one specific task, and I think that is the sensible future where on-device AI is heading. I’m sure we

15:05 will see a lot more examples of these micro AI models that are super efficient at one single task, so we can build interesting apps around them. But, what do you think about Desert Ant Labs and their models? Have you tried any of these? What are your first impressions? Let us know in the comment section down below. And folks, if you like these types of technical breakdowns, please let me know by smashing that like button underneath the video. And also, don’t forget to subscribe to our channel. This has been Andrus from Better Stack, and I

15:35 will see you in the next videos.