Desert Ant launches 18 on-device AI models, claiming 300x real-time transcription on iPhone

I got excited about a fast transcription model, but Voz appears to be Parakeet v3 with new macOS and iOS-specific inference code.

We optimized Parakeet for the Apple Neural Engine with our own inference path, reaching about 300x real-time on iPhone 16 and 17. Our next Voz is trained from scratch and should be at least twice as fast, with Android and other platforms coming later.

I would get more use from a model that turns PDFs into a JSON schema, or generates titles and tags for posts. I am mainly interested in web apps.

We have a Schemer model coming for free-form text to structured JSON. Image-to-JSON is next on our list after that.

These models could improve the CMS work I do for clients, but most look iOS-only and the benchmarks use recent iPhones. I doubt a cheap VPS gets comparable speed.

Only a few models are iOS-first because of sequencing. We plan to make all models cross-platform in the coming weeks.

This looks like proprietary packaging and marketing over open models: Voz is Parakeet 0.6B v3, Clear is DeepFilterNet 3, and Ear is Whisper-tiny's language predictor.

With frontier LLM help, a recent Mac, and a recent iPhone, porting an existing model to the ANE is becoming a fairly straightforward hill-climbing exercise.

Why is Core ML the only target so far? Android has comparable options such as ML Kit.

Most models are planned for Android and web, though Voz, Clips, and Title take longer to port properly. The issue is mostly release order.

I like local specialized models even more than local LLMs, but I do not see the business model. If I already have the weights and run inference offline, why should the vendor keep earning from my customers?

You are conflating price with cost. A product can be worth paying for even when the vendor does not pay an inference cost per request.

Charging only after an app exceeds 100,000 monthly active devices seems unusually well aligned. A company at that scale can pay for the software it relies on.

Subscriptions pay for continuing updates as hardware and local-model tooling change, and a monthly expense is often easier for a startup to approve than a large upfront purchase.

The vendor owns the IP and sets the license terms. Downloading weights does not mean unrestricted use, and a free tier for 100,000 devices is generous.

A per-model-version purchase might fit local models better: buy the current weights, then decide whether a later update is worth paying for.

I tried the Clear demo and could not hear any difference between the raw and enhanced audio. Either my ear is not trained enough or the demo is broken.

DeepFilterNet 3 is small and fast but not especially good. MossFormer2 is better for commercially usable denoising, while Nvidia RE-USE can also remove reverb but has a non-commercial license.

On-device hate-speech triage sounds risky. What happens when moderation is automated locally?

I read it as a specialized model for uses such as game lobbies or parental filtering, where an indie developer needs moderation without paying large cloud-model bills.