Every developer we spoke to at WWDC this week had the same story: a wish list of AI features they'd build if cost weren't a factor, or expensive tokens they spend on existing features they'd happily swap for a local model. AFM 3 Core is great for a lot of use cases, but many developers found that a general-purpose 3b param model isn’t always the best fit for highly specific tasks, where specialized models can be both faster and better.

Tools like MLX and Core AI enable lightning-fast inference on Apple Silicon, but we're missing the models and streamlined SDKs to actually take advantage of them. That's why we're building Desert Ant Labs, a European on-device AI lab. We're going to ship dozens of small, opinionated audio and visual models and SDKs that drop into any product with a few lines of code. No inference cost. Nothing leaving the device. Running on the 6 billion devices people already own.

The first models are already running in Detail and Subwave, saving us hundreds per month in API costs and lettings us build intelligent features without giving up privacy or security. When inference cost drops to zero and inference speed is 10x faster, product design changes at a fundamental level. https://desertant.ai