In short
- EmbeddingGemma 2 is Google DeepMind's open multimodal embedder on the Gemma 4 architecture: text, code, images, video and audio in one 768-dimensional vector space.
- Four sizes on Ollama — 270m to 740m, 304MB to 1.3GB — all runnable on laptops; only the 440m and 740m tags accept images.
- The big wins over v1 are code retrieval (~14% better, MTEB code 78.68 vs 68.76) and native multimodality; multilingual text quality is flat (61.36 vs 61.15).
- MRL truncation to 256d cuts vector storage 3x with minimal quality loss — 128d only for text-only workloads.
- Run it with
ollama run embeddinggemma-2, but remember it's an embedder for RAG and search, not a chat model, and its native input window is 8K tokens.
EmbeddingGemma 2 has landed on Ollama. It's an open multimodal embedding model from Google DeepMind, built on the Gemma 4 architecture, and it maps text (including code), images, video and audio — and combinations of them — into a single, unified 768-dimensional vector space. If you run local RAG or semantic search, this is the interesting part: one small model covering four modalities, sized for laptops and even phones.
It's not a chat model. It produces vectors, not text — you pair it with a generator model in your RAG pipeline and never talk to it directly.
What embeddinggemma-2 is
The full model has 740M parameters, but it's modular: a 270M-parameter text model (a 130M transformer backbone plus a 140M embedder), with a 170M vision encoder and a 300M audio encoder you can load selectively. Load only the modalities your use case needs, and skip the rest.
Google's stated target is consumer hardware — mobile devices and laptops — with low-latency, on-device semantic search, RAG, classification and clustering in mind. That's a sensible brief: embedding work is a single forward pass, no autoregressive generation, so small models go a long way here.
What's new versus EmbeddingGemma 1
The clearest way to see what changed is the benchmark table on the model page: EmbeddingGemma 1 has figures only in the text columns, while v2 adds image, video and audio benchmarks it simply has no predecessor for. As far as the published comparisons show, v1 handled text only. The headline changes:
- Native multimodality. Four modalities — text, images, video, audio — in one shared 768d embedding space. There are no predecessor figures for the vision and audio benchmark columns, because v1 didn't do those tasks.
- Code. MTEB (code, v1) NDCG@10 of 78.68 versus 68.76 for EmbeddingGemma 1 — roughly a 14% improvement on code retrieval. This is the standout gain.
- Multilinguality. 100+ languages. But multilingual MTEB v2 barely moved: 61.36 versus 61.15. If you're text-only and happy with v1, raw text quality is not the reason to upgrade.
- Matryoshka Representation Learning (MRL). The 768d output can be truncated natively to 512d, 256d or 128d and re-normalised — up to a 6x reduction in vector storage costs with, per the readme, minimal quality impact down to 256d.
- Task-steered representations. Lightweight text instruction prefixes tune embeddings for search, classification, clustering or semantic similarity.
- Context. An 8K-token native input window — the readme says that covers minutes of audio or video.
Sizes, memory and which one to pick
The rule of thumb for any local model: it needs roughly its file size again in RAM or VRAM, plus headroom for context. EmbeddingGemma 2 is unusually light by that measure.
| Tag | Download size | Inputs | Sensible hardware |
|---|---|---|---|
:270m | 304–378MB | Text | Almost anything — phones, old laptops, low-power boxes |
:440m | 671–714MB | Text, Image | Any modern laptop; integrated graphics or a small GPU is fine |
:570m | 918–990MB | Text | Laptop CPU is plenty |
:740m / latest | 1.3GB | Text, Image | Any gaming GPU, Mac, or modest homelab box |
Note the modality split: only the 440m and 740m tags accept images. The 270m and 570m are text-only builds. The readme lists audio and video as supported modalities of the full model, so check what your engine actually exposes per tag before building a mixed-media pipeline.
Even the largest variant is a 1.3GB download — small enough that no variant forces a hardware decision on you.
How to run it with Ollama
The default pull is the full 740m build:
ollama run embeddinggemma-2
That fetches 1.3GB and gives you text + image embeddings. Pin a size if you want something smaller:
ollama run embeddinggemma-2:270m # smallest, text only
ollama run embeddinggemma-2:440m # text + image, 671–714MB
Two practical notes:
- Context window. The Ollama page lists a 256K context window for all variants, but the readme gives the model's native input window as 8,192 tokens. Know the 8K figure before you point it at long documents — chunk your corpus accordingly.
- You don't chat with it. Embedding models are called through the API — Ollama's
/api/embeddingsendpoint, for instance — from your RAG pipeline or search code, not run interactively like a chat model.
Two settings worth using from day one: the task instruction prefixes (a short prefix per task, per the readme) and MRL truncation. Truncate to 256d and re-normalise: that's 1:3 compression on your vector store with, per the readme's own table, minimal quality impact — multilingual MTEB falls only from 61.36 to 60.41. For text-only workloads, 128d (1:6) is viable.
What it's good for — and its limits
Where it shines:
- Local RAG and semantic search over mixed text, code and image collections — one embedder, one vector space, no stitching separate models together.
- Code retrieval, the clearest gain over v1 (78.68 vs 68.76 on MTEB code).
- Multilingual corpora — 100+ languages, though quality is flat versus v1 rather than better.
- On-device and laptop deployments — the whole design brief.
- Cheap vector storage via MRL truncation.
Where it doesn't:
- Long documents. The 8K-token input window means you can't embed a whole long document in one pass. Chunk first.
- Aggressive truncation on multimodal work. The readme's truncation table shows the asymmetry clearly: at 128d, multilingual MTEB falls from 61.36 to 57.89, but MMEB overall drops from 59.01 to 45.65. Truncation hurts multimodal work much faster than text. Stick to 256d or above if images are in play.
- Generation. It's an embedder. Pair it with a chat model for RAG; don't expect it to answer anything.
- Benchmark caveat. All published results use the full-precision checkpoint. Your quantised local build may differ — as usual, test on your own data before committing.
How it compares with other local embedding models
The established text-only choices for local RAG are nomic-embed-text, bge-m3, mxbai-embed-large and Qwen3-Embedding — all available in Ollama and well-supported in the usual vector-database tooling. We're not going to claim a ranking among them: the readme doesn't compare against any of them, and we'd want independent, same-protocol numbers before doing so.
| Model | Modalities | Size class | Best use |
|---|---|---|---|
| embeddinggemma-2 (270m–740m) | Text, code, image, video, audio | 304MB–1.3GB | Mixed-media RAG; code retrieval; on-device |
| nomic-embed-text | Text | Small | General text RAG |
| bge-m3 | Text | Mid-sized | Multilingual text; established default |
| qwen3-embedding (0.6b) | Text | Small | Small text embeddings |
EmbeddingGemma 2's differentiator isn't raw text quality — multilingual MTEB is flat versus v1, and there are no published head-to-head figures against the rivals above. The differentiator is multimodality in one model and one vector space. If your corpus mixes text, images and audio, one embedder covers it.
For API users, cost won't decide it: on the marketplace, embeddinggemma-2:270m runs from £0.0058 per 1M tokens, against nomic-embed-text at £0.004 and bge-m3 at £0.01.
What we'd pick: text-only RAG on a laptop — the 270m tag, truncated to 256d. Text plus images — the 440m. If you're already running nomic-embed-text or bge-m3 and your corpus is text-only, there's no urgent reason to switch; switch when the mixed-media corpus arrives.
Worth hosting?
Embedding models are among the easiest things to serve on spare hardware: sub-1GB footprints, no autoregressive generation, and they slot neatly alongside chat models on a single mid-range GPU or a decent Mac — the 740m variant included, with room to spare.
If you'd rather call than host, the 270m and 440m variants are already live on the marketplace at £0.0058 and £0.0072 per 1M tokens. And if you have hardware sitting idle, hosting embedding models is a gentle way to start earning from it — see providers.
Questions people ask
How do I run embeddinggemma-2 in Ollama?
ollama run embeddinggemma-2 pulls the default 740m tag (1.3GB, text + image). Pin a size with ollama run embeddinggemma-2:270m for the smallest text-only build, or :440m if you need image embeddings. Call it through the API (e.g. /api/embeddings), not interactively.
Which embeddinggemma-2 size should I pick?
Text-only RAG: 270m. Text plus images: 440m. The 570m is text-only but larger than the 440m; the 740m is the full text + image build. Only 440m and 740m accept images.
How much RAM does embeddinggemma-2 need?
Roughly the download size again, plus headroom for context. The 270m is 304–378MB, the 440m 671–714MB, the 570m 918–990MB, and the 740m 1.3GB.
Can embeddinggemma-2 embed images and audio?
Yes — the full model covers text, images, video and audio in one 768d space. On Ollama, the 440m and 740m tags accept text and image inputs; audio and video are supported modalities of the full model per the readme.
What does Matryoshka (MRL) truncation mean for my vector database?
You can truncate the 768d output to 128d, 256d or 512d and re-normalise, cutting storage up to 6x. 256d (1:3) loses little quality; 128d (1:6) is best for text-only workloads and hurts multimodal scores much more.
Models in this post — live on Inpherio
Have a GPU, a Mac or a spare PC?
Let it earn while you sleep.
Run open models for people who need them. You set the price and keep everything but the 15% fee.
Become a hostBuilding with AI?
Every open model, one API.
OpenAI-compatible. Priced by the people who run the hardware.
Start free See the marketplaceRead “EmbeddingGemma 2 on Ollama: Google's multimodal embedder that runs on a laptop” at https://inpherio.co.uk/blog/embeddinggemma-2-on-ollama-google-s-multimodal-embedder-that-runs-on-a