Squeezebits

SqueezeBits is a Seoul, South Korea-based startup developing AI model lightweighting (compression) and inference optimization technology. Its core technique is quantization: reducing 32-bit model data to 4 bits or fewer while preserving model accuracy, which lowers the memory and compute required during inference and speeds up model execution. The company describes its work as building an ecosystem for designing, training and deploying highly compressed neural networks, and its technology is applied across mobile smartphones, laptops, edge devices and GPU cloud environments, supporting image, video, speech and natural-language models.

The company's products include the OwLite toolkit, which allows users who are not compression specialists to lighten, compare and analyze AI models, and Yetter, a generative AI API service built on its optimization inference engine for image and video generation with LLM services planned. SqueezeBits also publishes technical work on inference engines and hardware, including benchmarking of Intel Gaudi accelerators, guided decoding on vLLM and SGLang, disaggregated inference on Apple Silicon (NPU prefill with GPU decode), vocabulary trimming for small language models, and GraLoRA, a block-wise variant of LoRA fine-tuning. Its RoBoost Agent work integrates NVIDIA Cosmos world models to generate synthetic data for physical AI applications such as robotics and autonomous driving.

Founding story

The company was founded by researchers from the Neural Processing Unit (NPU) research team at POSTECH's graduate school. Its co-founders had published model-compression research at machine learning conferences including CVPR, NeurIPS and ICLR over the preceding seven years, with more than 70 international papers on deep learning acceleration, and had experience designing AI-specific hardware. Hyungjun Kim is a co-founder and chief executive officer. As of May 2024 Forbes described the company as founded two years earlier.

Business model

SqueezeBits sells AI optimization software, historically on a subscription basis, and also offers a generative AI API service (Yetter) built on its own inference engine. It has engaged in proof-of-concept and optimization projects with enterprise customers and in co-development partnerships with AI hardware vendors.

Subscription sales of optimization software, per Forbes, alongside a generative AI API service positioned as delivering image and video generation at reduced cost.

Traction

More than 20 companies, including Naver and SK Telecom, had completed technology verification (PoC) or projects with SqueezeBits as of January 2024. The company worked with global AI hardware firms including Intel and NVIDIA shortly after founding, and co-developed NPU model compression software with Rebellions in 2024.

Latest developments

In July 2026 Rebellions announced an agreement to acquire SqueezeBits as its first acquisition, following Rebellions' selection as the first direct investment target of Korea's National Growth Fund and its 2024 merger with Sapeon Korea; SqueezeBits is to continue operating independently. Recent product and research activity includes the Yetter generative AI API service, OwLite v2.5 with Qualcomm Neural Network support, Intel Gaudi optimization work, and RoBoost Agent synthetic data generation using NVIDIA Cosmos.