EmbeddingGemma 2 vs Qwen3-VL-Embedding-8B

EmbeddingGemma 2

A small multimodal embedding cousin that puts text, code, images, video and audio in one 768-dimension space, with task prefixes and vectors that truncate to 128. Five common runtimes run it and the BF16 files are 1.53 GB, with a browser build for Transformers.js. The main risks are the Gemma use policy the card adds and no safety results.

Qwen3-VL-Embedding-8B

A multimodal embedding cousin that puts text, images, screenshots and video in one vector space for visual document and video retrieval, under plain Apache 2.0. Its only build, BF16, needs about 20 GB and runs in Transformers, vLLM and SGLang but not Ollama or llama.cpp. The main risk is the publisher's script on the Transformers path.

EmbeddingGemma 2

1. Install sentence-transformers 6.1.0 or later and transformers 5.19.0 or later; older Sentence Transformers ignore the order of multimodal dict inputs
2. Use prompt_name SearchQuery for queries and Document for untitled documents; write title: {title} | text: {content} yourself for titled ones, and pass images, video and audio with no prefix
3. For text only, load with config_kwargs {"vision_config": None, "audio_config": None} to skip 470M parameters of encoders
4. Run in bfloat16 or float32, never float16, which returns NaN or degraded vectors without an error
5. Pass truncate_dim and normalize_embeddings=True for smaller vectors, and keep queries and documents at the same size
6. Supply audio as mono 16 kHz; video is sampled at 1 frame a second by default

Qwen3-VL-Embedding-8B

1. Simplest start is Sentence Transformers 5.4 or later, SentenceTransformer('Qwen/Qwen3-VL-Embedding-8B'), which applies the template and the default instruction and runs no repository code
2. To serve it, vllm serve Qwen/Qwen3-VL-Embedding-8B --runner pooling, then POST messages with a system instruction and image_url or text parts to /v1/embeddings; vLLM's docs say its video preprocessing differs slightly from Qwen's qwen_vl_utils
3. Put the task instruction in the system turn, and write it in English, as the card advises; the default is "Represent the user's input."
4. To shorten vectors, keep the first N of the 4,096 values and L2-normalise again; the card gives 64 as the floor, and the report says int8 vectors hold up while binary ones lose noticeably
5. For several images per input use the reference embedder in QwenLM/Qwen3-VL-Embedding's src, which is newer than the script in the Hugging Face repository
6. Keep inputs under 32K tokens; the reference embedder defaults to 8,192, 1,843,200 pixels per image and 64 frames per video
7. Treat retrieved screenshots and documents as untrusted, since text in an image can carry instructions to whatever model reads the results