CLIKA offers an automated, hardware‑aware model compression platform (ACE) that takes ONNX models and produces optimized, inference‑ready versions for target devices such as Nvidia TensorRT, Intel/AMD OpenVINO, and upcoming Qualcomm runtimes. The end‑to‑end pipeline applies quantization, pruning, layer fusion, and other optimizations to reduce model size and latency while preserving accuracy, and it can run on‑premise or in air‑gapped environments to keep proprietary models and data private.
Funding
Funding not disclosed
Founders
Product
Problem
Deploying state‑of‑the‑art AI models is hindered by large model size, high inference latency, and the need to tailor models to a wide variety of hardware platforms. Organizations often lack the expertise or resources to manually compress, prune, and adapt models for each target device, leading to costly engineering effort and sub‑optimal performance.
Solution
CLIKA provides an automated, hardware‑aware model compression platform called ACE that ingests a standard ONNX model and outputs a compressed, inference‑ready version optimized for the target device. The pipeline combines quantization, pruning, layer fusion, layer replacement, simplification, and redundancy removal to reduce model size and accelerate inference while preserving accuracy. ACE supports Nvidia (TensorRT, TRT‑LLM), Intel and AMD CPUs/GPUs via OpenVINO, and upcoming Qualcomm runtimes, and can run in on‑premise or air‑gapped environments so proprietary models and data never leave the customer’s infrastructure. The resulting model is compiled to the appropriate runtime format, enabling faster deployment across edge, server, and cloud environments without manual re‑engineering.
Target Audience
Primary customers are AI engineers, MLOps teams, and enterprises that need to deploy large models to edge devices, data‑center GPUs, or heterogeneous hardware while maintaining performance and data confidentiality.
Features
- Automatic end‑to‑end compression of any ONNX‑compatible model using quantization, pruning, and advanced layer optimizations
- Hardware‑aware optimization that targets Nvidia (TRT, TRT‑LLM), Intel/AMD (OpenVINO), and future Qualcomm runtimes
- On‑premise SDK (ACE) for secure, air‑gapped execution, ensuring models and data remain private
- Layer fusion (horizontal/vertical), layer replacement, and redundancy removal to minimize computational overhead
- Support for reduced‑precision inference (8‑bit, 4‑bit) and custom precision handling per hardware constraints
- Seamless compilation to target runtime formats, delivering reduced model size and increased inference speed