Product1 publisher3 min readPublished
SmartGym 8 routes its prompts to a private cloud its own developer built
Matt Abras says SmartGym is not eligible for Apple's Private Cloud Compute through the Foundation Models framework, so the app's natural-language workout builder calls frontier models on a private cloud he built and runs.
The Product Desk · Product desk
What happened
- SmartGym 8 shipped with AI workout creation and deeper Siri integration, built for the iOS 27, iPadOS 27 and watchOS 27 releases.
- The app uses Apple's new Language Model APIs to extend Foundation Models out to frontier models running on SmartGym's own private cloud, with requests kept anonymous and disconnected from user identity.
- For exercise questions, the app feeds exercise execution images into Apple's on-device Foundation Models to produce more descriptive answers about individual movements.
- Siri can now start SmartGym sessions and control an active one with commands such as "Next exercise in SmartGym" or "Log set in SmartGym."
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Private Cloud Compute eligibility now decides architecture for a small iOS team. A developer outside it has to stand up an inference backend before the cloud-model feature can ship at all.
- exposure The privacy guarantee is the developer's to defend. If a proxy is misconfigured or a log survives, the user's complaint arrives at SmartGym rather than at Apple.
- cost A one-developer app now carries the cost of running inference servers for every routine it generates, and 9to5Mac's write-up does not say what that costs or whether users pay for it.
- decision Teams building for iOS 27 have to pick a model location feature by feature, keeping what fits on the phone and funding infrastructure for anything that needs a larger model.
Somebody stands in a gym, taps the + button on SmartGym's Routines tab, and types "create a push-pull-legs workout with 10 exercises per day. Avoid bench press. Include 2 warm-up exercises per day." [8] When leg day lands wrong, they type "move leg day to Friday" and it moves [9]. Nothing on that screen tells the user which computer answered.
Two computers answer, depending on the feature. Exercise questions with images go into Apple's on-device Foundation Models [15]. The routine builder goes out to frontier models on infrastructure SmartGym operates, reached through Apple's new Language Model APIs, which let a developer extend Foundation Models code to a remote model [5]. The two inference locations have two different owners [16].
The reason for the split, according to 9to5Mac, is eligibility. Abras says SmartGym is currently not eligible to use Apple's Private Cloud Compute through the new Foundation Models framework, so he built his own [4]. 9to5Mac did not disclose why the app is ineligible [17].
What he took on with it is a promise. "Prompts and responses are never logged, cached, or retained by SmartGym," Abras wrote, describing the service as private by design and every request as anonymous and disconnected from the user's identity [6]. The app never connects directly to an AI lab, and by his account the labs behind those frontier models never receive prompts and do not train on what users submit [7]. A promise like that has an operator behind it. Someone runs the anonymising layer and answers the support email when a user asks where their prompt went.
The Siri half of the release rides on Apple's plumbing instead. "SmartGym now integrates more deeply with Siri, making routines, exercises, equipment, muscle groups, schedules, and other training information accessible through natural voice requests," Abras wrote [11]. Users can start a session by voice and say "Next exercise in SmartGym" mid-workout [12]. Asking which routines are scheduled this week returns visual results inside the Siri interface [13], and Siri's on-screen awareness handles "Open the third routine on the screen" [14].
For anyone shipping an iOS 27 feature this quarter, two things decide whether it costs a backend: whether the feature fits the on-device model, and whether it needs a retention promise a user could hold you to. On-device with no promise required is the cheap quadrant, and Apple's model is already on the phone. Bigger model plus a promise is the expensive one: hosting, anonymisation, and the wording support uses when a user asks what is kept. If you cannot name in one sentence who receives the prompt and what they keep, you need an inference backend you have not stood up.
Abras can name it. His app cannot use the framework that would have let Apple carry that guarantee [4], and on his account the labs behind SmartGym's models never see a prompt [7].
What to watch
- Whether Apple opens Private Cloud Compute to third-party apps through the Foundation Models framework, and on what eligibility terms.
- Whether SmartGym publishes pricing or a subscription tier that covers the cost of its own cloud inference.
- Whether Abras names the frontier model providers behind SmartGym Intelligence or publishes an audit of the no-retention claim.