Mobilint designs the NPU architecture, compiler, quantizer and runtime together.
SDK qb is the software stack for every Mobilint NPU. qb Compiler turns trained models into NPU executables, and qb Runtime runs them in your application through the Mobilint device driver.
High performance per watt within edge power limits.
More of the run spent computing, less spent waiting on data.
490+ validated models, from CNNs to LLMs and VLMs.
One core design, from a single core to multi-chip systems.
qb Compiler converts trained models into MXQ, the format Mobilint NPUs run. One compiler covers ARIES and REGULUS alike.
Models built and validated on a GPU carry over as they are, from vision models to LLMs.
INPUT
PyTorch, ONNX, TensorFlow and more.
01
Converts your model into one common graph.
02
Sets precision layer by layer, using sample inputs.
03
Builds an MXQ file for the target NPU.
OUTPUT
Loads with qb Runtime on ARIES and REGULUS.
FRAMEWORKS
Of FP16 accuracy retained with mixed-precision quantization, with no retraining.
Validated model architectures, from CNNs to LLMs and VLMs.
The first two rows are Top-1 accuracy; the YOLO rows are AP.
WikiText2 perplexity (PPL); lower is better.
Browse a growing list of latest pre-compiled models with pre- and post-processing included.
qb Runtime loads compiled MXQ models and schedules them on the NPU cores. Split one request across cores for lower latency, or give each core its own request for higher throughput.
01
Bring your model
Compile TensorFlow, PyTorch or ONNX models to MXQ with qb Compiler.
02
Load and request model
Load the MXQ file and send inference requests.
03
Choose the acceleration cores
Create an NPU instance and pick which cores run the model.
04
Call from your app
Python and C++ APIs on Linux and Windows.
Python
C++
Linux
Windows
GCC
MSVC
All cores work on one input. Use it for real-time, single-batch work.
Cores work on separate inputs. Use it for multi-channel video and batch workloads.
