Hồ Việt Anh

Published writing

RTIC editorial cover for a benchmark of 17 Ollama embedding models across nine tasks, with document-to-vector illustration.

research-note · 2026-10-26

A 300M model ties Qwen3-8B: benchmarking 17 Ollama embedding models

We ran 17 embedding models from Ollama's catalogue through nine public tasks covering English and Vietnamese RAG, code and short-text tasks. Four models tie at the top with an aggregate of about 0.73, but their runtime differs sevenfold. Supplementary tests cover Vietnamese legal fine-tuning and visual document retrieval.

Project contributions

No public project contribution has been linked yet.