Sionic AI, the company behind the STORM Platform(an enterprise AI agent builder), has announced comsat-embed, a family of retrieval embedding models developed under its EUREKA (Extremely Universal Robust Embedding for Knowledge Access) project.
The results set a new bar for Asian-language retrieval. On the Korean MTEB Retrieval v2 benchmark, comsat-embed-ko-8b ranked #1 among all open-weight models with an average NDCG@10 of 0.7930, ahead of models more than three times its size. On the Japanese JMTEB v2 benchmark, comsat-embed-ja-8b took the overall #1 spot with 0.8133. Critically, comsat-embed also outperformed leading commercial embedding APIs from Google, Cohere, and OpenAI in both languages, with its advantage confirmed on out-of-distribution benchmarks spanning 17 industry domains including legal, healthcare, finance, and manufacturing.
The performance stems from a training pipeline designed around real-world search conditions: diverse document curation, synthetic query generation, explicit control of answer-position bias, positive-aware hard negative mining, and KL-divergence-based knowledge distillation.
comsat-embed powers the Agentic RAG engine inside Sionic AI's STORM Platform, delivering high-accuracy enterprise search for customers across banking, manufacturing, and finance. Preview versions are available on HuggingFace under CC BY-NC 4.0, with commercial licensing available directly from Sionic AI.
Read the full technical deep-dive at blog.sionic.ai/comsat-embed.