Stella
NovaSearchStella is a family of English text-embedding models from the NovaSearch team, built by further training Alibaba's gte-large-en-v1.5 and gte-Qwen2-1.5B-instruct. The 400M and 1.5B checkpoints were trained with Matryoshka representation learning, so each can emit vectors from 512 up to 8,192 dimensions, and the publishers recommend inputs of up to 512 tokens. The 1.5B model later served as the starting point for NovaSearch's Jasper embedder.
Openness
3 high confidence- weights
- open (1.5B and 400M checkpoints, ungated on Hugging Face)
- data
- closed (the embedding training data is neither released nor named, and the GTE backbones' own pretraining belongs to Alibaba)
- code
- partial (RAG-Retrieval publishes the stage-3 training code shared by Stella and Jasper, not the full recipe)
- license
- MIT (OSI)
Stella's weights are published under the MIT license, and part of its training code is public. The data it was trained on has not been released or listed, so the training cannot be repeated.
- https://huggingface.co/api/models/NovaSearch/stella_en_1.5B_v5 recorded 2026-11-12
Model API for NovaSearch/stella_en_1.5B_v5: `gated: false`, cardData.license `mit`, with model.safetensors among the files.
- https://huggingface.co/api/models/NovaSearch/stella_en_400M_v5 recorded 2026-11-12
Model API for NovaSearch/stella_en_400M_v5: `gated: false`, cardData.license `mit`.
- https://huggingface.co/NovaSearch/stella_en_1.5B_v5/raw/main/README.md recorded 2026-11-12
Model card: "The models are trained based on `Alibaba-NLP/gte-large-en-v1.5` and `Alibaba-NLP/gte-Qwen2-1.5B-instruct`"; FAQ: "The training method and datasets will be released in the future"; front matter `license: mit`.
- https://raw.githubusercontent.com/NovaSearch-Team/RAG-Retrieval/master/README.md recorded 2026-11-12
RAG-Retrieval README: "RAG-Retrieval released the core training code (stage3) of Stella and Jasper embedding model".
Adoption
2 high confidenceHugging Face downloads summed across the 1.5B and 400M checkpoints, most of them for the 400M model.
- https://huggingface.co/api/models/NovaSearch/stella_en_1.5B_v5 recorded 2026-11-12
"downloads":16961 for NovaSearch/stella_en_1.5B_v5, the trailing 30-day count the Hub API reports.
- https://huggingface.co/api/models/NovaSearch/stella_en_400M_v5 recorded 2026-11-12
"downloads":56702 for NovaSearch/stella_en_400M_v5, the trailing 30-day count the Hub API reports.
Capability
4 medium confidenceStella's 1.5B model ranked among the strongest English embedders on MTEB at the end of 2024, level with GTE's best instruct model on the same English average. It covers English only and is tuned for short inputs, which keeps it behind the multilingual, long-context leaders.
- https://arxiv.org/html/2412.19048 recorded 2026-11-12
Report Table 1, MTEB results as of December 24, 2024: Stella_en_1.5B_v5 average (56 datasets) 71.19, NV-Embed-v2 72.31, bge-en-icl 71.67, SFR-Embedding-2_R 70.31, gte-Qwen2-1.5B-instruct 67.16.
- https://huggingface.co/NovaSearch/stella_en_1.5B_v5/raw/main/README.md recorded 2026-11-12
Model card FAQ: "512 is recommended" for sequence length, "the model is trained on datasets of 512 length"; the model-index lists 56 MTEB English results.
Verified 2026-11-12, except some scores, still awaiting re-confirmation