New Giga-Embeddings-instruct embedding model family released openly

🤖 GigaChat opens Giga-Embeddings-instruct embedding models

The GigaChat team (ai-sage organization on Hugging Face) released Giga-Embeddings-instruct embedding models in three sizes — 480M, 3B, and the flagship 10B-A1.8B — all under the MIT license. The models convert text into vectors for RAG and semantic search; 480M and 3B are built on Qwen3, while the flagship is based on DeepSeek-V3 (MoE, ~1.8B active parameters).

🌍 Open Russian-language embeddings without reliance on paid APIs — a foundation for local RAG stacks and enterprise search. According to the model cards, 10B-A1.8B scores 74.99 on MTEB (Russian), 3B raised MTEB (code) from 62.37 to 76.93, and 480M outperformed FRIDA-820M at half the size.

👤 Downloadable on Hugging Face and runnable via sentence-transformers, vLLM, or SGLang (vLLM currently requires a build from main).

Source 1: https://huggingface.co/ai-sage/Giga-Embeddings-instruct-480M-0826

Source 2: https://huggingface.co/ai-sage/Giga-Embeddings-instruct-3B-0826