Giga-Embeddings
Sber's MIT-licensed Russian and English text embedding models in 480M, 3B and 10B mixture of experts sizes.
In short. Giga-Embeddings turns Russian and English text into number vectors that a search system can compare, so it can find passages that mean the same thing as a question. It comes in a small 480M version, a 3B one and a 10B mixture of experts, all free under the MIT license.
Summary
Giga-Embeddings-instruct is the text embedding line from Sber's GigaChat team, tuned for Russian and English. The 0826 generation has three models. The 480M model uses a Qwen3-style architecture with 28 layers and 1,024-dimension vectors, the 3B model has 36 layers and 2,048-dimension vectors, and the 10B model is a DeepSeek-V3-style mixture of experts with 1.8B active parameters, multi-head latent attention and 64 routed experts, 4 of them used per token, plus one shared. All three turn the attention of a language model bidirectional, use mean pooling followed by L2 normalization, and were trained with a contrastive InfoNCE objective. The 3B model first appeared on Hugging Face on 29 July 2026 and the other two on 10 August 2026. Queries need a one-sentence task instruction in the form "Instruct: ... Query: ...", while documents are encoded as they are, and CLS or last-token pooling gives wrong results.
In Sber's tables the 480M model scores 70.98 on MTEB Russian, against 63.64 for Qwen3-Embedding-0.6B, 64.02 for EmbeddingGemma-300M and 70.95 for FRIDA-820M. Qwen3-Embedding-0.6B is ahead on English (70.70 against 69.52), code (75.41 against 72.87) and multilingual (64.33 against 56.97). The 3B model scores 74.57 on Russian, 71.93 on English, 76.93 on code and 63.9 on multilingual, up from 74.16, 71.07, 62.37 and 55.51 for the earlier 3B model, and the 10B mixture of experts scores 74.99, 72.23, 78.40 and 65.60. The card also lists throughput on an H100 at batch size 16, with the 10B model at 112.6k per second at 512 tokens, against 87.9k for the 3B, 70.1k for Qwen3 Embedding 4B and 42.6k for Nemotron 8B, without stating the unit. All of these numbers are the developer's own.
The license is MIT, and the BF16 files are 0.98 GB, 6.31 GB and 20.96 GB. The repositories ship custom bidirectional modeling code, so loading needs trust_remote_code, and the cards show Sentence Transformers and vLLM examples. The configuration files list 8,192 positions for the 480M and 3B models and 262,144 for the 10B model, but the cards state no context length. In the 30 days to 11 November 2026, Hugging Face counted about 22,000 downloads for the 480M model, 2,900 for the 3B and 1,100 for the 10B.
Good for
- Russian and English semantic search and retrieval-augmented generation, with a permissive MIT license.
- A small 480M embedder that is close to a much larger Russian-language model on Sber's tables.
- Code retrieval and multilingual tasks with the 3B and 10B models, which score higher than the old 3B model.
- Running in vLLM at high throughput, which the card highlights for the 10B mixture of experts.
Watch out for
- Queries need an "Instruct: ... Query: ..." prefix and documents must not have one, and mean pooling with L2 normalization is required.
- The cards document Russian and English only, and the 480M model trails Qwen3-Embedding-0.6B on English, code and multilingual tests.
- All scores are Sber's own, and the throughput table does not state its unit.
- The code needs trust_remote_code, and the 10B model is a 21 GB file even though only 1.8B parameters are active.
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Giga-Embeddings. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| 480M (1,024-dimension vectors) BF16 | 480M | 0.98 GB | 1.13 GB | Laptop or phone |
| 3B (2,048-dimension vectors) BF16 | 3.15B | 6.31 GB | 7.26 GB | Laptop or phone |
| 10B-A1.8B mixture of experts BF16 | 10.5B 1.8B active | 21 GB | 24 GB | Gaming GPU |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
- Best open-source embedding models in 2026
- Sber's open models: a 432B chat model, speech and embeddings
Frequently asked questions
- What is Giga-Embeddings?
- Giga-Embeddings turns Russian and English text into number vectors that a search system can compare, so it can find passages that mean the same thing as a question. It comes in a small 480M version, a 3B one and a 10B mixture of experts, all free under the MIT license.
- Can I use Giga-Embeddings commercially?
- Yes. Giga-Embeddings is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Giga-Embeddings open source?
- Giga-Embeddings is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Giga-Embeddings on my own computer?
- Yes. The smallest variant, 480M (1,024-dimension vectors), needs about 1.13 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Giga-Embeddings released?
- Giga-Embeddings was released by Sber (GigaChat) on July 29, 2026.
- How long a context window does Giga-Embeddings have?
- Giga-Embeddings supports a context window of up to 8K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Giga-Embeddings-instruct-480M-0826 model card (Hugging Face) huggingface.co
- Giga-Embeddings-instruct-3B-0826 model card (Hugging Face) huggingface.co
- Giga-Embeddings-instruct-10B-A1.8B-0826 model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Harrier OSS v1 Microsoft | 270M – 27B | 32K | MIT (Permissive) | 0.621 GB Laptop or phone | |
| Qwen3-Embedding Alibaba Qwen | 600M – 7.57B | 32K | Apache 2.0 (Permissive) | 1.38 GB Laptop or phone | |
| EmbeddingGemma 2 Google DeepMind | 270M – 740M | 8K | Apache 2.0 (Permissive) | 0.621 GB Laptop or phone | |
| BGE-M3 BAAI | 568M | 8K | MIT (Permissive) | 2.61 GB Laptop or phone |