Curiosity · AI Model
GTE-Qwen2 7B Instruct
GTE-Qwen2-7B-Instruct is a 7-billion-parameter open text-embedding model from Alibaba's DAMO Academy, released July 2024. It uses the Qwen 2 7B language backbone, an instruction-aware query template, and a 4096-dimensional dense embedding head, topping the MTEB retrieval leaderboard in both English and Chinese at release. It supports 32k context, is released under Apache 2.0 on Hugging Face, and has become a common production embedding choice for enterprise RAG.
Model specs
- Vendor
- Alibaba DAMO Academy
- Family
- GTE
- Released
- 2024-07
- Context window
- 32,768 tokens
- Modalities
- text
Strengths
- Topped MTEB and C-MTEB at release
- Open Apache 2.0 licence
- 32k context — good for long passages
- Instruction-aware query formatting
Limitations
- Heavy at 7B — larger than most production embedders
- 4096-dim vectors increase storage cost
- Inference latency higher than smaller BGE / E5 models
- Requires GPU for reasonable throughput
Use cases
- Enterprise RAG over long documents
- Multilingual semantic search (English + Chinese + more)
- Dense retrieval for recommendation systems
- Clustering and deduplication at scale
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MTEB English (avg) | ≈70.2 (#1 at release) | 2024-07 |
| C-MTEB Chinese | ≈72.1 (#1 at release) | 2024-07 |
Frequently asked questions
What is GTE-Qwen2-7B-Instruct?
A 7B open text-embedding model from Alibaba DAMO, based on Qwen 2, released July 2024 with top MTEB leaderboard scores.
Is it multilingual?
Yes. It inherits Qwen 2's multilingual training and leads both MTEB (English) and C-MTEB (Chinese) at release.
Is the licence commercial-friendly?
Yes — released under Apache 2.0 on Hugging Face.
Sources
- GTE-Qwen2-7B-Instruct on Hugging Face — accessed 2026-04-20
- MTEB leaderboard — accessed 2026-04-20