Dnotitia

Dnotitia is an AI and data infrastructure company established in 2023. Its work centers on "vector intelligence": converting multimodal data into searchable vector representations and serving retrieval and inference workloads at low cost. The company's two flagship products are Seahorse, a vector database intended for large-scale multimodal semantic search and positioned around speed, accuracy, cost efficiency, filtering and automated tuning, and Mnemos, an edge AI device described as a personal LLM device that runs models locally without data-center or cloud infrastructure through LLM optimization and compression, low-power operation and automatic tuning and deployment.

Alongside the hardware and database layers, Dnotitia develops small language models (SLMs) and retrieval-augmented generation (RAG) systems, and offers an AI Console focused on secure operation and data privacy. Its GenAI assistant is built on the in-house DNA (Dnotitia AI) foundation model family, which the company positions for Korean-language and Korean-cultural understanding with English support. Dnotitia distributes models publicly through a Hugging Face organization with 48 models and 2 datasets listed, including DNA 3.0 variants (such as a 397B-parameter DNA3.0-397B-A17B-FP8 release and image-text-to-text models ranging from roughly 1B to 123B parameters), patched Qwen3 chat-template releases, and a research paper on making Qwen3 reason in Korean using reinforcement learning. It also operates a public leaderboard site benchmarking major LLMs on Korean.

Business model

Dnotitia sells AI infrastructure products: the Seahorse vector database (including a Seahorse Cloud service offered with free starter credits), the Mnemos edge LLM device, and an AI Console for private/secure AI deployment. It also publishes open model weights and datasets on Hugging Face.

Traction

The company cites more than 100 industry experts working with it, 9 or more major government LLM development projects won, and 34 or more strategic partnerships across industries. Its Hugging Face organization lists 48 models, 2 datasets, 8 collections and 34 team members, with datasets of roughly 10M and 9.55M rows and model downloads in the tens to hundreds per listing.

Latest developments

Recent model publishing activity includes DNA-VL-STEER-2B and DNA3.0-397B-A17B-FP8, and the company promotes Seahorse Cloud for building AI agents from customer data as well as a leaderboard comparing major LLMs in Korean.