GTE-Qwen2-7B-Instruct
Alibaba's 7B instruction-tuned embedder that ranked #1 on MTEB in both English and Chinese
About
GTE-Qwen2-7B-Instruct builds on the Qwen2-7B decoder with bidirectional attention grafted in, producing 3584-dimensional embeddings over a 32K-token window. Released June 16, 2024, it ranked first on both the English and Chinese MTEB evaluations — a rare bilingual double — and Alibaba reports an MTEB average of 70.24. It uses instruction prefixes for asymmetric retrieval (queries and documents get different prompts). The practical catches: it needs trust_remote_code and bfloat16 to load correctly (the MTEB team documented silent quality collapse on wrong configs), and like all 7B embedders it is GPU-hungry. Alibaba also sells a separate commercial embedding API, so the open weights and the paid service are not the same model.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
MTEB average 70.24; C-MTEB 72.05 (Alibaba-reported)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: June 16, 2024 release date verified in xorbitsai integration issue; 3584 dims, 32K context, Apache 2.0 corroborated across MTEB loader docs and model directory pages.; MTEB 70.24 / C-MTEB 72.05 are Alibaba-reported figures — labeled as such in the entry.; MTEB team's documented loader pitfalls (bf16, trust_remote_code) add credibility to the technical details..