granite-embedding-97m-multilingual-r2
by ibm-granite
Highest-quality multilingual embedding under 100M parameters for edge and mobile deployment
ibm-granite/granite-embedding-97m-multilingual-r2mixpeek://text_extractor@v1/ibm_granite_embed_97m_multi_r2Overview
Granite Embedding 97M is IBM's ultra-lightweight multilingual text embedding model that achieves the best retrieval quality of any open model under 100M parameters. Using ModernBERT architecture with model pruning and vocabulary selection, it produces 384-dimensional embeddings across 200+ languages while fitting in under 100MB quantized.
On Mixpeek, this model fills the edge deployment gap — embedding generation on devices, in browser workers, or on minimal hardware where the 300M+ models in the catalog are too large. It supports 32K token context, making it suitable for embedding longer documents without chunking.
Architecture
ModernBERT with model pruning and vocabulary selection. 97M parameters. 384-dim output embeddings. 32K max context. Trained on 52 languages + code with enhanced multilingual data. ONNX quantized weights are 98MB.
Mixpeek SDK Integration
import { Mixpeek } from "mixpeek";const mx = new Mixpeek({ apiKey: "API_KEY" });await mx.collections.ingest({collection_id: "multilingual-docs",source: { url: "https://example.com/documents.json" },feature_extractors: [{feature: "text_embeddings",model: "ibm-granite/granite-embedding-97m-multilingual-r2"}]});
Capabilities
- Best retrieval quality under 100M params (59.6 nDCG@10 on MMTEB)
- 200+ language support with 52-language enhanced training
- 384-dimensional embeddings
- 32K token context window
- 98MB quantized — runs on edge devices and in-browser
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| MMTEB Retrieval (18 tasks) | nDCG@10 | 59.6% | IBM Research, 2026 — arxiv:2605.13521 |
Performance
Specification
Research Paper
Granite Embedding Multilingual R2 Models
arxiv.orgBuild a pipeline with granite-embedding-97m-multilingual-r2
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Open Studio