What is the Quartz inference engine?
Quartz is Saient’s native inference runtime. It is responsible for turning supported model files into useful on-device output without handing execution to a hosted API or a third-party inference runtime.
In Saient mobile, Quartz powers local GGUF language-model sessions and native SD1.5 image generation. The application owns the runtime; downloaded model weights remain independent files governed by their own licences.
How Quartz runs a model locally
- Validate the model. Quartz checks the expected format, structure and tensor metadata before execution begins.
- Prepare native inputs. Tokenization and input preparation happen locally. The runtime maps model data without converting it through ONNX.
- Execute Saient-owned operators. Quartz performs the tensor operations, scheduling and memory coordination required by the supported model path.
- Dispatch image work through Vulkan. The mobile image pipeline uses Saient’s Vulkan compute path to run the supported SD1.5 workload on the phone GPU.
- Write the result on-device. Text responses and generated images are returned directly to Saient without a cloud inference round trip.
What Saient owns—and what it does not
Saient-owned runtime
Model parsing, supported tokenization, tensor execution, scheduling, memory coordination and Vulkan dispatch live in Quartz.
Separate model weights
GGUF language models and SD1.5 weights are downloaded separately and retain their original open-weight licences.
No cloud inference
The network is used for model delivery when needed. Inference itself runs on the user’s device.
No ONNX Runtime. No inference API.
Quartz does not include Microsoft ONNX Runtime, and the signed Saient mobile application does not bundle an ONNX model or ONNX runtime library. The supported inference path is native to Saient.
After a model is downloaded, Quartz can run supported chat and image workflows without sending prompts, model tensors or generated images to an inference service.
Enough detail to explain the system—not reproduce it
This page describes the product boundary and execution stages so users can understand what runs on their hardware. Saient does not publish Quartz’s low-level kernel implementation, shader layouts, memory-planning heuristics, operator-fusion choices or optimisation formulas.
That distinction keeps the privacy and ownership claims verifiable without turning the engine’s implementation into a public blueprint.
Quartz engine questions
Does Quartz use ONNX Runtime?
No. Quartz uses Saient’s native inference implementation for its supported paths.
Does Quartz need the internet?
A connection is needed to download models from Saient or another model source. Once installed, supported inference runs locally.
Does Quartz own the model weights?
No. Quartz is the runtime. Model weights remain separate assets under their own licences.
Run supported AI models through a runtime built for Saient.
Start with Saient, download a model that fits your hardware, and keep inference on your device.