Embedding models - verified
Open-weight embedding models you can run yourself
10 embedding models publish their weights on Hugging Face, so you can index documents without sending them to an API. They range from 149M parameters (GTE ModernBERT Base) to 8.38B (Nemotron 3.5 Embed 8B). 10 of 10 ship under licenses that allow commercial use.
Comparing hosted APIs too? All embedding models with list prices. Looking for chat models to run locally? Open-weight LLMs by memory.
- Models
- 10
- Commercial use
- 10
- Flexible dims
- 8
- Beyond text
- 2
Sorted by size
Largest by total parameters first; a model whose card states no count sits at the end. Size is not a quality ranking. Each model page has the official usage snippet.
| Model | Params | Default dims | Max input | Inputs | License | Commercial |
|---|---|---|---|---|---|---|
| Nemotron 3.5 Embed 8BNVIDIAnvidia/Nemotron-3.5-Embed-8B-BF16 | 8.38B | 4,096Flexible | 32,768 | Text, Image | openmdw-1.1 | Commercial use allowed |
| Qwen3-Embedding-8BQwen (Alibaba)Qwen/Qwen3-Embedding-8B | 7.57B | 4,096Flexible | 32,768 | Text, Code | apache-2.0 | Commercial use allowed |
| Qwen3-Embedding-4BQwen (Alibaba)Qwen/Qwen3-Embedding-4B | 4.02B | 2,560Flexible | 32,768 | Text, Code | apache-2.0 | Commercial use allowed |
| EmbeddingGemma 2Google DeepMindgoogle/embeddinggemma-2 | 744M | 768Flexible | 8,192 | Text, Code, Image, Audio, Video | apache-2.0 | Commercial use allowed |
| Qwen3-Embedding-0.6BQwen (Alibaba)Qwen/Qwen3-Embedding-0.6B | 596M | 1,024Flexible | 32,768 | Text, Code | apache-2.0 | Commercial use allowed |
| Nomic Embed Text v2 MoENomic AInomic-ai/nomic-embed-text-v2-moe | 475M | 768Flexible | 512 | Text | apache-2.0 | Commercial use allowed |
| Voyage 4 NanoVoyage AIvoyageai/voyage-4-nano | 346M | 1,024Flexible | 32,000 | Text | apache-2.0 | Commercial use allowed |
| Granite Embedding 311M Multilingual R2IBMibm-granite/granite-embedding-311m-multilingual-r2 | 312M | 768Flexible | 32,768 | Text, Code | apache-2.0 | Commercial use allowed |
| GTE ModernBERT BaseAlibaba (Tongyi Lab)Alibaba-NLP/gte-modernbert-base | 149M | 768 | 8,192 | Text | apache-2.0 | Commercial use allowed |
| BGE-M3BAAIBAAI/bge-m3 | Not stated | 1,024 | 8,192 | Text | mit | Commercial use allowed |
Method
How to read these numbers
Read from official provider docs, pricing pages and Hugging Face model cards on . Prices and limits change; check the linked source before you commit to re-embedding a corpus.
- First-party list prices
- Prices are the provider's own list price per 1M input tokens at the standard rate, with the batch rate when one is listed, copied from its pricing page. Discounts, free tiers and cloud marketplace prices are not included.
- Not a quality benchmark
- Nothing here measures retrieval quality or speed. Tables sort by price, input length or dimensions only. Shortlist on these facts, then evaluate on your own queries and documents.
- Not stated means not stated
- When the provider's docs do not give a value, the cell says so instead of guessing. Some providers price per character, image or request; those prices stay in the note on the model page.
- Snippets copied verbatim
- Usage snippets come from the page linked under each one: the provider's docs or the Hugging Face model card. Models whose sources publish no snippet show none.
FAQ
What is the smallest open-weight embedding model here?
GTE ModernBERT Base, at 149M parameters. Small models embed faster on a CPU or a laptop GPU; size is not a quality ranking.
Which open-weight embedding models can I use commercially?
10 of 10: Nemotron 3.5 Embed 8B, Qwen3-Embedding-8B, Qwen3-Embedding-4B, EmbeddingGemma 2, Qwen3-Embedding-0.6B, Nomic Embed Text v2 MoE, Voyage 4 Nano, Granite Embedding 311M Multilingual R2, GTE ModernBERT Base and BGE-M3. Each model page links the license text; read it before you ship.
Which open-weight embedding model takes the longest input?
Nemotron 3.5 Embed 8B, Qwen3-Embedding-8B, Qwen3-Embedding-4B, Qwen3-Embedding-0.6B and Granite Embedding 311M Multilingual R2, at 32,768 tokens per input.
Which open-weight embedding model outputs the most dimensions?
Nemotron 3.5 Embed 8B and Qwen3-Embedding-8B, at up to 4,096 dimensions.
Can an open-weight model embed images or audio?
Yes: Nemotron 3.5 Embed 8B (text and image) and EmbeddingGemma 2 (text, code, image, audio and video).
How do I run them?
Each model page shows the usage snippet from its Hugging Face card or the provider's docs, most often a few lines of Sentence Transformers. For API prices and the hosted models, see https://www.developersdigest.tech/models/embeddings.
Newsletter
New open weights, every week
One email per week. New model releases, open-weight drops and what they mean for what you build.
Keep exploring
Model API pricing and specs
First-party prices per 1M tokens for chat and reasoning models.
Local models
Open-weight LLMs you can download and run, with sizes and licenses.
What is RAG?
Where embeddings fit in a retrieval pipeline, end to end.
Vector databases compared
Where the vectors go once you have picked a model.