Evidence: benchmark results vendor-reported
EmbeddingGemma 2 — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
A multimodal embedding model built on Gemma 4: a 270M text backbone plus optional vision (170M) and audio (300M) encoders that you load only when you need them. It returns 768-dimension vectors that can be truncated to 512, 256 or 128 dimensions, and reads up to 8,192 tokens of input. Ollama carries it as embeddinggemma-2, in 270m, 440m, 570m and 740m sizes.
EmbeddingGemma 2 needs about 2 GB of VRAM at BF16 — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
Get EmbeddingGemma 2 running
The cheapest catalogued GPU that runs EmbeddingGemma 2 is the Intel Arc B570 (10 GB).
or compare on Vast.ai from $0.35/hr (typical low · varies)
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run EmbeddingGemma 2
Ollama carries it as embeddinggemma-2 (270m text-only, 440m with images, 570m with audio, 740m with everything). Pull it with `ollama pull embeddinggemma-2:270m` and call it through the embeddings API (ollama.embed); an embedding model is not chatted with, so there is no `ollama run` command.
Specifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 740 Million (270M text + 170M vision + 300M audio)
- Context window
- 8,192
- Architecture
- Gemma 4-based embedding encoder, 768-d output (Matryoshka 512 / 256 / 128)
- Provider
- Google DeepMind
- Licence
- Apache-2.0
- Availability
- Weights downloadable
- Practical deployment
- Edge device (phone, SBC)
- Specified at
- BF16
- System RAM
- 4 GB
- Record updated
- 2026-11-21
Commercial use permitted. No usage restrictions beyond attribution.
Embedding output
- Vector size
- 768
- Can be shortened to
- 512, 256, 128
- Input limit
- 8,192 tokens
- Inputs
- text, code, images, video, audio
- Load only what you need
- Text only: 270M · Text + images: 440M · Text + audio: 570M · All inputs: 740M
An embedding model turns content into vectors for search and RAG. It writes no text, so this page has no speed estimate.
Limits and caveats
- Run it in bfloat16 or float32. In float16 its numbers overflow and it returns invalid or silently degraded embeddings without an error (Google's card).
- Embeddings from different models do not mix: switching from EmbeddingGemma to EmbeddingGemma 2 means re-indexing your documents.
- Benchmarks on the card are vendor-reported.
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
| Benchmark | Score | Provenance |
|---|---|---|
| MTEB multilingual v2 (768d) | 61.36 / 100 | vendor-claimed · https://huggingface.co/google/embeddinggemma-2 |
| MTEB English v2 (768d) | 68.46 / 100 | vendor-claimed · https://huggingface.co/google/embeddinggemma-2 |
| MTEB code v1 (768d) | 78.68 / 100 | vendor-claimed · https://huggingface.co/google/embeddinggemma-2 |
EmbeddingGemma 2 — frequently asked questions
How much VRAM does EmbeddingGemma 2 need?
About 2 GB at BF16 — quantized weights plus framework overhead, before any KV cache. We publish no context-length figure or speed for it: it is an embedding model: it turns text, images and audio into vectors and writes no text, so there is no decode speed or growing KV cache to model. Apple Silicon counts unified memory toward the same figure.
Does EmbeddingGemma 2 run on an RTX 4090 (24 GB)?
Its weights and overhead come to about 2 GB at BF16, but we do not publish a fit verdict or a speed for it: it is an embedding model: it turns text, images and audio into vectors and writes no text, so there is no decode speed or growing KV cache to model.
How do I run EmbeddingGemma 2 locally?
Ollama carries it as embeddinggemma-2 (270m text-only, 440m with images, 570m with audio, 740m with everything). Pull it with `ollama pull embeddinggemma-2:270m` and call it through the embeddings API (ollama.embed); an embedding model is not chatted with, so there is no `ollama run` command. Running the published tag would send your prompts to a hosted GPU rather than your own machine.