Tailored to the model.
Tuned to the machine.
Built for agents.
Sharar tailors its weights, kernels, memory, and decoding to each model family and each device it supports, so many agents run fast on one desktop machine. NVIDIA DGX Spark today; Mac and AMD Strix Halo coming soon.
The model
Weights, drafters, and state chosen for each model family.
The machine
Kernels and memory plans for its bandwidth, memory, and storage.
The agents
Many agents at once, each resuming where it left off.
Quick start
Pick your device and model, then run the command. Sharar runs on the NVIDIA DGX Spark today.
- Needs DGX OS, NVIDIA driver 580 or newer, and the CUDA 13 toolkit.
- uv installs Python 3.12 and the pinned PyTorch.
- The first serve downloads the model and builds the GB10 kernels once.
Device
Model
Profile
Sparks
Run this command