product updateGoogle DeepMind

Unsloth publishes GGUF quantizations of Google DeepMind's 740M-parameter EmbeddingGemma 2 multimodal embedding model

TL;DR

Unsloth has released GGUF quantizations of EmbeddingGemma 2, Google DeepMind's open 740M-parameter embedding model for text, images, video and audio. The base model maps all four modalities into one 768-dimensional space with an 8K context window under Apache 2.0.

3 min read
0

Unsloth has published GGUF quantizations of EmbeddingGemma 2, Google DeepMind's open-weight embedding model that maps text, images, video and audio into a single 768-dimensional vector space. The upload is a community repackaging for local runtimes. The underlying model is by Google DeepMind and is licensed under Apache 2.0.

What Unsloth released

The repository, unsloth/embeddinggemma-2-GGUF, holds quantized versions of Google's checkpoint. Unsloth claims its Dynamic 3.0 quantization method "achieves superior accuracy & outperforms other leading quants." That is a company claim, and the source gives no quantization-specific benchmarks. Quant levels and file sizes are not stated in the source material.

All benchmark figures below come from Google's full-precision checkpoint, not the GGUF builds.

EmbeddingGemma 2 specifications

  • Parameters: 740M total. This is a 270M text model (130M transformer plus 140M embedder), a 170M vision encoder and a 300M audio encoder.
  • Architecture: 24 layers, model dimension 512, hidden dimension 2048, 4 attention heads, GQA/MQA attention, 1024-token sliding window, 5:1 local-to-global ratio, 262,144-token vocabulary, mean pooling, and a 512→768 projection layer.
  • Context window: 8,192 tokens. Google says this is enough for minutes of audio or video.
  • Output: 768 dimensions natively. Matryoshka Representation Learning (MRL) supports truncation to 512, 256 and 128 dimensions.
  • Languages: 100+. Google also says code tasks improve about 14% over the predecessor.
  • Base: Built on Gemma 4 architecture advances, per Google.
  • Pricing: Not applicable for local weights. No hosted API pricing was disclosed. The training cutoff date was not disclosed.

The vision and audio encoders load selectively. Effective size is 270M for text only, 440M for text and image, 570M for text and audio, and 740M for full multimodal.

Reported benchmarks (768d, full precision)

Benchmark EmbeddingGemma 2 EmbeddingGemma 1
MTEB multilingual v2 61.36 61.15
MTEB code v1 (NDCG@10) 78.68 68.76
MIEB lite 64.64 n/a
MMEB v2 Image (Hit@1) 57.28 n/a
MMEB v2 VisDoc (NDCG@5) 67.84 n/a
MMEB v2 Video (Hit@1) 50.67 n/a
MSEB Retrieval (MRR@10) 69.54 n/a
MAEB 49.39 n/a

Text quality is nearly flat against the first-generation model on multilingual MTEB (+0.21). The gain is on code retrieval (+9.92 points). The image, video and audio benchmarks have no first-generation baseline because the original model was text-only.

Truncation trade-offs

Google's truncation table shows how quality degrades as vectors shrink:

  • 512d: MTEB multilingual 61.17, MMEB v2 overall 58.38
  • 256d: 60.41 and 56.24
  • 128d (1:6 compression): 57.89 and 45.65

Google says quality impact is minimal down to 256d and that 128d suits text-only workloads. Multimodal quality drops sharply at 128d: MMEB v2 falls from 59.01 to 45.65, and MSEB retrieval from 69.54 to 56.71.

Usage requirements

The model uses task instruction prefixes on text inputs only, such as task: search result | query: {query} for search queries and title: {title} | text: {content} for documents. Omitting prefixes still works but reduces precision, per Google. Images, video and audio take no prefix. Truncated vectors must be re-normalized (L2) before cosine similarity, or ranking degrades silently. Selective encoder loading differs across libraries, so GGUF runtime behavior should be checked against each tool's documentation.

What this means

A single embedding space for four modalities in under 1B parameters, with an Apache 2.0 license, lowers the barrier to on-device multimodal retrieval. The selectable 270M text-only footprint keeps it practical for phones and laptops. The GGUF release matters mainly for the llama.cpp ecosystem and similar local runtimes, though multimodal support in those tools may lag text-only support. That has not been verified here.

The open questions are quantization loss and tooling support. Embedding models are sensitive to precision because small vector shifts change nearest-neighbor rankings. Until Unsloth or third parties publish MTEB or MMEB results for the quantized files, treat Google's numbers as an upper bound. Teams should also measure retrieval quality on their own data at the quant level they plan to deploy.

Comments

Loading...