GTE (Alibaba) - General Text Embeddings
Basic Information
- Product Name: GTE (General Text Embeddings)
- Developer: Alibaba (Alibaba NLP)
- Country/Region: China
- Hugging Face: https://huggingface.co/collections/Alibaba-NLP/gte-models
- Type: Open-source general text embedding model series
- License: Apache-2.0
- Latest Models: gte-multilingual-base, gte-Qwen2 series
Product Description
GTE (General Text Embeddings) is a series of general text embedding models developed by Alibaba NLP, specifically designed for NLP tasks such as semantic search, RAG, and text re-ranking. The GTE series offers various model sizes ranging from large (8B parameters) to small (0.3B parameters), supports 70+ languages, and employs an encoder-only architecture for efficient inference. The latest version supports elastic embedding and a context window of 8192 tokens.
Core Features
- Multilingual Support: Supports 70+ languages
- 8192 Token Context: GTE v1.5 series supports input up to 8192 tokens
- Elastic Embedding: Adjusts output dimensions on demand, significantly reducing storage costs
- Encoder-only Architecture: 10x faster inference compared to decoder-based LLM architectures
- Multiple Sizes: Various parameter sizes including 8B, 2B, 1.5B, and 0.3B
- gte-Qwen2 Series: Instruction-following embedding models based on the Qwen2 architecture
- gte-multilingual-base: Latest multilingual general embedding model
Model Matrix
| Model | Parameters | Architecture | Features |
|---|---|---|---|
| gte-Qwen2-7B-instruct | ~8B | Decoder | Highest accuracy, supports instructions |
| gte-Qwen2-1.5B-instruct | ~2B | Decoder | Balanced performance |
| gte-multilingual-base | ~0.3B | Encoder | Multilingual, efficient inference |
| gte-large-en-v1.5 | ~0.4B | Encoder | English-specific, 8192 context |
| gte-base-en-v1.5 | ~0.1B | Encoder | Lightweight English version |
Business Model
- Completely Open Source and Free: Apache-2.0 license
- Alibaba Cloud Integration: Available through Alibaba Cloud AI services
- Hugging Face Availability: Directly downloadable and usable from Hugging Face
Target Users
- Application developers needing multilingual embeddings
- RAG systems pursuing high inference efficiency
- AI application developers in the Alibaba Cloud ecosystem
- Teams needing elastic embedding to reduce costs
- Chinese and multilingual search scenarios
Competitive Advantages
- Encoder architecture offers 10x faster inference than decoder architectures
- Elastic embedding reduces storage costs
- Multiple sizes for flexible adaptation to different scenarios
- Alibaba's NLP technology expertise
- Strong capabilities in Chinese and multilingual contexts
Relationship with OpenClaw Ecosystem
The GTE series provides OpenClaw with flexible embedding model options. The efficient inference characteristics of the encoder architecture are suitable for OpenClaw's latency-sensitive real-time scenarios. The elastic embedding feature allows OpenClaw to dynamically adjust vector dimensions based on different storage budgets. The various sizes of GTE models also enable OpenClaw to choose the appropriate embedding model based on device performance.
External References
Learn more from these authoritative sources: