The GigaChat team (ai-sage organization on Hugging Face) has released a new generation of embedding models, Giga-Embeddings-instruct, in three sizes — 480M, 3B, and the MoE flagship 10B-A1.8B — all under the MIT license. The models convert text into vectors for RAG and semantic search and are already available for download on Hugging Face.

image
image
image

What happened

The models in the lineup use bidirectional (encoder-style) attention and contrastive learning with the InfoNCE loss. The 480M and 3B models are built on the Qwen3 architecture, while the flagship 10B-A1.8B is based on DeepSeek-V3 with MLA and MoE: 64 experts, 4 active per token, approximately 1.8 billion active parameters out of 10 billion total. According to the model cards, the 10B-A1.8B scores 74.99 on MTEB (Russian) and achieves up to 114.5k tokens per second on H100 with vLLM, outperforming Qwen3 Embedding 4B, Nemotron 8B, and F2LLM-v2-8B. The 3B model improved MTEB (code) from 62.37 to 76.93 — a gain of 14.56 points — while the 480M scored 70.98 on MTEB (rus, v1.1), outperforming the larger FRIDA-820M.

Context

The release is not a new architecture: the base models Qwen3 and DeepSeek-V3 are public, and the team adapted and fine-tuned them for the embedding task; the novelty lies in the training recipe, including distillation from an earlier version of the model, rather than in the architecture itself. Embedding models are a key component of RAG systems and semantic search: they represent text as vectors, allowing search in a corpus by meaning rather than by keywords. Prior to this release, top Russian-language embeddings were mostly tied to closed or paid solutions, while open options were weaker. The sparse MoE attention used by the flagship is a common way to achieve the quality of a large model at the inference cost of a compact one.

Why this matters for the industry

For the industry, these are the first open, top-tier Russian-language embedding models under the MIT license that can be integrated directly into a product without a paid API or vendor lock-in. For product teams, this is a ready-made critical component for local RAG stacks and enterprise search in three price tiers. The MoE flagship delivers 10B-model quality at the inference cost of approximately 1.8 billion active parameters, making dense search over large corpora significantly cheaper. Once support in vLLM and SGLang lands in stable releases, the lineup could become the default open Russian-language embedding for RAG, forcing competitors, including paid APIs, to justify their pricing.

Why this matters for users

All three models are already available for download on Hugging Face: the 480M and 3B can be run via sentence-transformers, which is a stable path, while the 10B-A1.8B can be run via vLLM from the main or nightly build, or via SGLang: PR #52948 has already been merged into the vLLM main branch but has not yet made it into a release, and SGLang support is coming via PR #35531. The 480M is suitable for compact Russian search, the 3B for code-related tasks, and the flagship 10B-A1.8B for maximum quality and multilingual support. Existing RAG systems should run an A/B comparison with their current embeddings before deployment.

What is still unknown / limitations

The evaluation base is mixed and self-reported in the model cards: the 480M was evaluated on MTEB (rus, v1.1), while the benchmark version for the flagship 10B-A1.8B is not specified. The 14.56-point gain for the 3B on MTEB (code) is a comparison with the team's own previous generation, not with competitors, and the claim of first place on ruMTEB is taken from the model card and requires verification against the leaderboard. The speed comparison of the flagship with Qwen3 Embedding 4B, Nemotron 8B, and F2LLM-v2-8B does not disclose measurement conditions — batch size, sequence length, quantization — and it is difficult to correctly compare an MoE model with 1.8 billion active parameters to dense 4B and 8B models in terms of throughput.

Sources

Author

Look at AI, editorial team