Family 01
Large Language Models
Llama, Mistral, Qwen, Phi — instruction-tuned and base checkpoints up to ~13 B parameters running at conversational latency.
Invotet SDK
Compile once. Deploy to every Invotet module.
A unified Python SDK that ingests PyTorch, ONNX, and HuggingFace checkpoints, quantizes for Invotet modules, and ships a deterministic runtime to the device. No CUDA in the loop.
Frameworks
Compile any of the three; ship one runtime
Framework
PyTorch
Trace or torch.export checkpoints compile directly with no rewrite.
Framework
ONNX
Standards-based interchange — compile any ONNX-exported model.
Framework
HuggingFace
transformers checkpoints land on Invotet through a one-line loader.
Five lines, one runtime
Compile the model. Connect the device. Generate.
Loaded from PyTorch, ONNX, or HuggingFace, your checkpoint quantizes for the module and runs through a deterministic on-device runtime over USB or PCIe.
invotet-sdk · v0.1.0
from invotet import Module, load_hf
# Compile a HuggingFace checkpoint for Invotet modules
model = load_hf("meta-llama/Llama-3-8B-Instruct", quantization="int8")
# Open the on-device runtime over USB or PCIe
with Module.connect() as device:
device.deploy(model)
response = device.generate(
"Plan a four-waypoint patrol around this perimeter.",
max_tokens=256,
)
print(response)Models supported
Frontier models, on the device that has to think for itself.
Family 02
Vision-Language Models
LLaVA, Qwen-VL, and Gemma-Vision class models — image + text reasoning on-platform without a downlink.
Family 03
Perception
YOLO, DETR, SAM, and custom detectors. Run side-by-side with an LLM in the same module.
Request early access.
The SDK is in early access while the toolchain matures. Tell us your framework, the checkpoint family, and the target module — we will get you provisioned.