Platforms & requirements
Slab 0.1 packages, device requirements, engine support, and installed sizes.
These requirements cover Slab 0.1.1 (10 November 2026). Platform minimums are unchanged from 0.1.0. Minimums are build targets; tested devices are listed separately.
At a glance
| iOS | Android | macOS library | macOS tools | |
|---|---|---|---|---|
| For | iPhone and iPad apps | Android apps | Mac apps, Python | ML engineers preparing models |
| Package | Swift package: Slab.xcframework + SlabSwift |
slab.aar |
libslab.dylib, slab.h, python/slab |
slab-load command-line tool |
| Minimum OS | iOS / iPadOS 18 | Android 12 (API 31) | macOS 15 | macOS 15 |
| Architecture | arm64 (devices); arm64 simulator | arm64-v8a | arm64 (Apple silicon) | arm64 (Apple silicon) |
| Size | 2.8 MB framework per slice (8.2 MB xcframework with its three slices) | 2.4 MB .aar; libslab.so 6.1 MB installed |
2.7 MB | 3.0 MB tool + 6.1 MB Android runner |
| Engines | CPU, GPU (Metal), Neural Engine | CPU, GPU (Vulkan), Qualcomm Hexagon NPU | CPU, GPU (Metal), Neural Engine | the same, for checking and benchmarking |
| APIs | Swift, C | Java / Kotlin, C (NDK) | C, Python | command line |
| Tested on | iPad Pro (M1), iPadOS 27.0 | Galaxy S23 Ultra (Android 17), Galaxy Z Fold7 (Android 16), Galaxy A07 (Helio G99, Mali-G57) | MacBook Pro (M1 Max), macOS 26.6 | the same Mac |
Not supported: Intel Macs, Windows, Linux, 32-bit or x86 Android, iOS before 18, Android before 12.
What every platform shares
- One shared library, nothing else to install. It links only the operating system's own libraries and
frameworks; no third-party runtime, no Python or ONNX Runtime on the device. Only the public C API
(
slab_*functions) is exported; everything else is hidden and stripped. No static libraries are shipped. - Models are
.slabpacks, converted on a Mac withslab-load convertfrom ONNX (opsets 10 to 20), or downloaded ready-made from the model gallery. One pack runs on every platform. - Precision. The CPU and the GPU compute in FP32; the Neural Engine and the Hexagon NPU in FP16. Packs store weights in FP32, FP16, INT8 or 4-bit. FP16 weights use half the storage of FP32. The release reports identical CPU/GPU outputs for an FP16 pack compared with the same weights widened to FP32; quantizing original FP32 weights may round values. See Prepare a model.
- Engine choice. The app picks CPU, GPU or NPU, or leaves it on automatic: the pack author's default, then the NPU, the GPU and the CPU, using what the device has.
- Runtime license. A license file names the app IDs that may load Slab (one primary ID and up to 15 variants of the same product; free 30-day trial
from the customer console). The same file works on iOS, Android and macOS and with
slab-load. - Protected models (encrypted packs) run on the CPU and the GPU, not on the Neural Engine or the NPU.
iOS and iPadOS
- Add to the app: the Swift package (Xcode 16 or later; Swift tools 6.0). The C header is inside the
framework as
Slab.h. - Devices: every iPhone and iPad that runs iOS 18 (A12 Bionic or later), all of which have a Neural Engine. The simulator slice is arm64 (Apple silicon Macs) and is meant for development.
- System frameworks used: Security, CoreFoundation, Foundation, Accelerate, Metal, MetalPerformanceShaders, MetalPerformanceShadersGraph; libz, libc++, libobjc.
- Model files: in the app bundle or in
Documents/. - License identity: the app's bundle identifier; each extension that loads Slab (widget, share, notification service) has its own and must be listed.
Android
- Add to the app:
slab.aar(Java and Kotlin API indev.slab,libslab.sofor arm64-v8a, R8 rules included). C and C++ code can callslab.hagainst the same library. - Minimum: Android 12 (API 31), 64-bit ARM. The library is aligned for 16 KB memory pages, for devices using 16 KB memory pages.
- System libraries used: libvulkan, libicu, libz, liblog, libandroid, libm, libdl, libc.
- CPU: ARMv8-A with NEON; dot-product and int8 matrix instructions are used when the CPU has them.
- GPU: Vulkan 1.1 with a compute queue, storage buffers of 64 MB or more, 128-byte push constants and host-visible memory. 16-bit storage is used when present. Tested on Adreno 740, Adreno 830 and Mali-G57.
- NPU: Qualcomm Snapdragon with a Hexagon HTP. Qualcomm's QNN libraries are not included: the app
bundles them from the Qualcomm AI Runtime SDK (
libQnnHtp.so,libQnnHtpPrepare.so,libQnnSystem.so, the HTP stub for the chip and its skel library), under Qualcomm's license. Tested with QAIRT 2.45 on Hexagon v73 (Snapdragon 8 Gen 2) and v79 (Snapdragon 8 Elite). Without them, Slab uses the GPU and the CPU. The NPU takes models with a single float32 input. - Model files: in the app's private
files/directory. - License identity: the
applicationId, including anyapplicationIdSuffix; callSlab.bind(context)before attaching the license.
macOS
Library. libslab.dylib with slab.h for Mac apps, and the Python library (python/slab, plain
Python 3 with ctypes, no compiled extension; tested with Python 3.11), which finds libslab.dylib where the
download puts it, or at SLAB_LIBRARY. The license checks the app's bundle identifier: an app bundle has one; a command-line program
embeds an Info.plist (-sectcreate __TEXT __info_plist), as the C setup guide shows.
Tools (slab-load). For ML engineers preparing models, not for shipping inside apps. The tool runs on your Mac; check, info, ops, bench and precision can also target a USB-connected Android phone with --device android (requires adb and USB debugging). Keep the bundled Android runner in libexec/slab-load/android-arm64/.
| Command | What it does |
|---|---|
convert |
ONNX to a .slab pack (--dry-run lists anything it cannot convert) |
info, check, ops |
What a pack contains; engine support and, with a license, actual pack loading; the operator table |
bench |
Load time, latency, session and peak memory, output comparison, and sustained frame rate with --duration |
quantize |
FP16, INT8 or 4-bit weights |
precision |
The precision each engine runs a pack at |
protect |
Encrypt a pack |
run |
Run a pack on raw input files and save the outputs |
Sizes, explained
- iOS: the framework binary for one architecture is 2.8 MB; an App Store build installs only the device slice. The xcframework download holds three slices (iOS device, iOS simulator, macOS).
- Android: the
.aaris 2.4 MB compressed; the native library is 6.1 MB on the device. Qualcomm's NPU libraries, when an app adds them, are extra. - macOS tools: the 0.1.1 bundle downloads as 4.96 MB and includes both
slab-loadand its Android runner. These tools are separate from the runtime library you ship in your app. - Models are separate: a pack's size is about its weights (for example MobileNetV2 is 15 MB as FP32, 8 MB as FP16).
Check your model
On your Mac, use the tools in the macOS bundle:
bin/slab-load convert model.onnx --dry-run
bin/slab-load check model.slab
bin/slab-load info model.slab
The dry run lists unsupported ONNX nodes before writing a pack. check reports which engines on this Mac can run the pack and its estimated memory. Use bin/slab-load check model.slab --device android to check on your connected Android phone; a Mac result does not certify Android support. With a license set, check also loads the pack on each compatible engine. See Prepare a model for the full workflow.
Operator support
Produced by slab-load ops --markdown on a Mac and on a Galaxy S23 Ultra. The converter also lowers these
ONNX operators into the ones below: Cast (value-preserving), Dropout, Identity, Reshape, Flatten, Squeeze,
Unsqueeze, ReduceL2, LpNormalization, IsNaN, QuantizeLinear, DequantizeLinear, DynamicQuantizeLinear,
ConvInteger and MatMulInteger (computed in float), and anything computed only from constants (Shape,
ConstantOfShape, Range and the like are folded at conversion).
| Operator | Apple CPU | Metal | Neural Engine | Android CPU | Vulkan | Hexagon NPU |
|---|---|---|---|---|---|---|
| Sub | yes | yes | yes | yes | yes | yes |
| Div | yes | yes | yes | yes | yes | yes |
| Add | yes | yes | yes | yes | yes | yes |
| Mul | yes | yes | yes | yes | yes | yes |
| Conv | yes | yes | yes | yes | yes | yes |
| Relu | yes | yes | yes | yes | yes | yes |
| MaxPool | yes | yes | yes | yes | yes | yes |
| GlobalAveragePool | yes | yes | yes | yes | yes | yes |
| Flatten | yes | yes | yes | yes | yes | yes |
| Gemm | yes | yes | yes | yes | yes | yes |
| Sigmoid | yes | yes | yes | yes | yes | yes |
| HardSigmoid | yes | yes | yes | yes | yes | yes |
| HardSwish | yes | yes | yes | yes | yes | yes |
| Transpose | yes | yes | yes | yes | yes | yes |
| LayerNormalization | yes | yes | yes | yes | yes | yes |
| MatMul | yes | yes | yes | yes | yes | yes |
| Erf | yes | yes | yes | yes | yes | yes |
| PRelu | yes | yes | yes | yes | yes | yes |
| BatchNormalization | yes | yes | yes | yes | yes | yes |
| Concat | yes | yes | yes | yes | yes | yes |
| Identity | yes | yes | yes | yes | yes | yes |
| InstanceNormalization | yes | yes | yes | yes | yes | yes |
| AveragePool | yes | yes | yes | yes | yes | yes |
| Shape | ||||||
| Reshape | ||||||
| LeakyRelu | yes | yes | yes | yes | yes | yes |
| Resize | yes | yes | yes | yes | yes | yes |
| ConvTranspose | yes | yes | yes | yes | yes | yes |
| Softmax | yes | yes | yes | yes | yes | yes |
| Clip | yes | yes | yes | yes | yes | yes |
| ArgMax | yes | yes | yes | yes | yes | yes |
| Gelu | yes | yes | yes | yes | yes | yes |
| Split | yes | yes | yes | yes | yes | yes |
| Gather | yes | yes | yes | yes | yes | |
| Unsqueeze | ||||||
| Slice | yes | yes | yes | yes | yes | yes |
| Squeeze | ||||||
| Pow | yes | |||||
| Sqrt | yes | yes | yes | yes | yes | yes |
| ReduceMean | yes | yes | yes | yes | yes | yes |
| Equal | yes | yes | yes | yes | yes | |
| Where | yes | yes | yes | yes | yes | |
| Expand | yes | yes | yes | yes | yes | yes |
| DynamicQuantizeLinear | ||||||
| Exp | yes | yes | yes | yes | yes | yes |
| ConstantOfShape | ||||||
| ReduceSum | yes | yes | yes | yes | yes | yes |
| Tanh | yes | yes | yes | yes | yes | yes |
| DequantizeLinear | ||||||
| Range | ||||||
| Less | yes | yes | yes | yes | yes | |
| Cos | yes | yes | yes | yes | yes | yes |
| Sin | yes | yes | yes | yes | yes | yes |
| DepthToSpace | yes | yes | yes | yes | yes | |
| Pad | yes | yes | yes | yes | yes | |
| DeformUnfold | yes | yes | yes | yes | ||
| DetrTopK | yes | yes | yes | yes | ||
| MaskToBoxLogit | yes | yes | yes | yes | ||
| Log | yes | yes | yes | yes | yes | yes |
| GridSample | yes | yes | yes | yes | yes | |
| Abs | yes | yes | yes | yes | yes | yes |
| Neg | yes | yes | yes | yes | yes | yes |
| Floor | yes | yes | yes | yes | yes | yes |
| Ceil | yes | yes | yes | yes | yes | yes |
| Round | yes | yes | yes | yes | yes | yes |
| Reciprocal | yes | yes | yes | yes | yes | |
| Elu | yes | yes | yes | |||
| Selu | yes | yes | yes | |||
| Max | yes | yes | yes | yes | yes | yes |
| Min | yes | yes | yes | yes | yes | yes |
| ReduceMax | yes | yes | yes | yes | yes | yes |
| ReduceMin | yes | yes | yes | yes | yes | yes |
| ReduceProd | yes | yes | yes | yes | yes | |
| CumSum | yes | yes | yes | yes | yes | |
| TopK | yes | yes | yes | yes | yes | |
| GatherND | yes | yes | ||||
| ScatterND | yes | yes | yes | yes | yes | yes |
| NonMaxSuppression | yes | yes | ||||
| SpaceToDepth | yes | yes | yes | yes | yes | |
| Mod | yes | yes | yes | yes | yes | |
| Greater | yes | yes | yes | yes | yes | |
| Not | yes | yes | yes | yes | yes | |
| And | yes | yes | yes | yes | yes | |
| Or | yes | yes | yes | yes | yes | |
| YoloNms | yes | yes | yes | yes | yes | yes |
| Cast | yes | yes | yes | yes | yes | |
| Resize(cubic) | yes | yes | ||||
| GatherElements | yes | yes | yes | yes | yes | yes |
| ReduceL2 | ||||||
| Trilu | ||||||
| RMSNormalization | yes | yes | yes | yes | ||
| Acos | yes | yes | yes | yes | yes |
A blank cell means there is no direct runtime implementation in this table. Some operations are lowered or folded by the converter as described above. Use convert --dry-run and check for your actual model.