GTE-Qwen2 VRAM requirements
Alibaba's multilingual text embedding models built on Qwen2, at 7 and 1.5 billion parameters, for search and retrieval over inputs of up to 32,000 tokens. Alibaba's figures run from 6.62 GB for GTE-Qwen2-1.5B-instruct to about 32 GB for GTE-Qwen2-7B-instruct, served with Infinity. The cheapest single GPU for the first is the 16 GB Tesla V100 at Verda (DataCrunch) ($0.26 an hour); for the second, the 48 GB RTX A6000 at Thunder Compute ($0.35).
What Alibaba says each setting needs
| Model and setting | Stated GPU memory | Cheapest single GPU with that much | Per hour | |
|---|---|---|---|---|
| GTE-Qwen2-7B-instructWeights at fp32, the model card's memory column | 26.45 GB | RTX A6000, 48 GB, Thunder Compute | $0.35 | Rent → |
| GTE-Qwen2-1.5B-instructWeights at fp32, the model card's memory column | 6.62 GB | Tesla V100, 16 GB, Verda (DataCrunch) | $0.26 | Rent → |
| GTE-Qwen2-7B-instruct, served with InfinityThe card's own serving command: bfloat16, batches of 8, on a GPU of compute capability 8.0 or later (Ampere onwards). It says about 16 to 32 GB, so the card shown holds the upper figure | about 32 GB | RTX A6000, 48 GB, Thunder Compute | $0.35 | Rent → |
The memory figures are Alibaba's, not our estimate. The price is the lowest on-demand rate for one GPU with at least that much memory, from our latest check; a card that holds a model is not a promise of its speed.
Before you rent
The two fp32 figures are the weights alone, before any batch of text is processed. Most servers run embeddings in bfloat16, which halves the weights, so the Infinity row, the card's own serving figure, is the one to rent for when the 7B model answers search queries; it also asks for an Ampere or newer card. Its smaller English models, gte-large-en-v1.5 and gte-base-en-v1.5, take 1.74 GB and 0.51 GB at fp32 and fit any GPU listed here.
In Alibaba's words
- GTE-Qwen2-7B-instruct: “| GTE-Qwen2-7B-instruct | Multilingual | 32000 | 3584 | 26.45GB |”
- GTE-Qwen2-1.5B-instruct: “| GTE-Qwen2-1.5B-instruct | Multilingual | 32000 | 1536 | 6.62GB |”
- GTE-Qwen2-7B-instruct, served with Infinity: “# requires ~16-32GB VRAM NVIDIA Compute Capability >= 8.0”
Source: the GTE-Qwen2-7B-instruct model card on Hugging Face, read 17 October 2026.
Text models are sized by a formula instead: LLM VRAM requirements for every open language model, or the LLM VRAM calculator for your own context and batch.
More models sized by their publishers' figures
- Whisper (speech recognition, OpenAI)
- Stable Diffusion 3.5 Medium (image generation, Stability AI)
- Wan2.1 (video generation, Wan-AI)
- Wan2.2 (video generation, Wan-AI)
- HunyuanVideo (video generation, Tencent)
- olmOCR (OCR and PDF conversion, Ai2)
- MinerU (OCR and document parsing, OpenDataLab)