ClikaRT — One AI Runtime for CPUs, GPUs, NPUs & TPUs

Heterogeneous execution

Different vendors. Different rules. One process.

CPUs, GPUs and NPUs each come with their own memory model, their own sync primitives, their own way of counting time. ClikaRT drives them all from a single process, each device behind its own async queue, all of them busy.

One graph · many devices

The code is the placement.

Tell a tensor where to live, and its computation follows. The runtime splits the graph, runs each side where you sent it, and stitches the results back together.

Models Formats

Bring your models as they are.

Load safetensors from HuggingFace, GGUF, or Diffusers pipelines directly, no conversion step, no intermediate format. ClikaRT detects what you give it and picks the right loader.

Compressed before it runs.

ACE, our Auto Compression Engine, quantizes and restructures your model before it reaches the runtime: up to 90% smaller, 18x faster, less than 1% accuracy loss. A model that needed a data-center GPU now fits on a phone, and ClikaRT runs it there.

Discover ACE

The infra layer between your AI models and device fleet.