Caramella Inc.

Better AI starts with better data and better systems.

Caramella is a South Korea-based AI engineering company working at the intersection of data-centric AI, language models, and practical AI systems. We turn unstructured information into usable intelligence and build the infrastructure needed to make AI reliable and efficient.

Aster: Efficient Language Model Research

Aster is our ongoing research and engineering initiative focused on building efficient language-model systems, from training infrastructure to inference.

Our areas of exploration include:

• Efficient architectures: state-space models (SSMs), recurrent computation, and hybrid attention approaches.
• Training infrastructure: Rust-based components, data pipelines, and reproducible experiments.
• Resource-efficient learning: mixed precision, memory-aware training, and model adaptation.
• Practical inference: quantization, serving efficiency, and evaluation across hardware environments.

We favor measurable improvements, clear trade-offs, and reproducible results over unverified performance claims.

What to Expect Here

This organization is the home for Caramella’s AI research artifacts. As projects become ready for public release, we aim to share selected models, datasets, benchmarks, and demos, subject to evaluation, licensing, and data-use requirements.

Individual releases will document their intended use, limitations, and evaluation evidence.

Connect

• Website: caramella.kr
• GitHub: github.com/caramellateam
• Hugging Face: huggingface.co/CaramellaInc

Caramella Inc. · South Korea