Sessions runs open language and image models entirely on-device. Bigger models are sharper but need more memory — so the catalog is grouped by how much RAM a device has. Sessions recommends models that fit yours.
The lightest models — fast, compact, and runnable on virtually every supported iPhone and iPad. The Ternary Bonsai models run entirely on CPU: no GPU needed, so they work on any supported device.
The everyday sweet spot — capable mid-size models for recent iPhone and iPad. Where most people start.
Larger, more capable models for iPhone Pro, M-series iPad, and Mac — including a reasoning model.
The heaviest, highest-fidelity models — Mac only. iOS can’t host models this large, so the Max tier runs on Apple silicon Macs. The Bonsai 27B builds are the outlier: 1-bit and ternary quantization shrink a 27-billion-parameter model to 3.8–7.2 GB, with custom Metal kernels that run GPU-accelerated on Apple silicon.
Tiers are guidance based on available memory, not a fixed device list — Sessions highlights the models that fit your hardware and lets you download any that do. Sizes are approximate downloads; every model runs entirely on-device. Models are sourced from the in-app catalog (powered by Hugging Face).