granite-embedding-311m-multilingual-r2

by ibm-granite

200+ language embedding model with 32K context and ModernBERT architecture

156Kdl/month

311Mparams

HuggingFace Use in Pipeline

Identifiers

Model ID

ibm-granite/granite-embedding-311m-multilingual-r2

Feature URI

mixpeek://text_extractor@v1/ibm_granite_embed_311m_multi_r2

Overview

Granite Embedding 311M Multilingual R2 is IBM's second-generation multilingual text embedding model, built on ModernBERT with alternating attention mechanisms and GeGLU activations. It achieves a 13-point improvement over R1 on MTEB Multilingual Retrieval (65.2) while supporting 200+ languages, 9 programming languages, and a 32K token context window.

On Mixpeek, this model excels at cross-lingual retrieval across global document collections. Its 32K context handles full-length legal contracts, research papers, and technical documentation without chunking. The Apache 2.0 license and broad deployment options (ONNX, OpenVINO INT8, vLLM, GGUF) make it suitable for production at scale.

Architecture

ModernBERT backbone with 22 layers, 12 attention heads, alternating attention patterns, and GeGLU activations. 311M parameters. Rotary position embeddings (RoPE) supporting 32K context. Trained via knowledge distillation from multiple teachers with contrastive fine-tuning and model merging. Matryoshka representation learning for flexible output dimensions.

Mixpeek SDK Integration

from mixpeek import Mixpeek

mixpeek = Mixpeek(api_key="YOUR_API_KEY")

mixpeek.ingest.documents(
    collection="global_contracts",
    source={"type": "s3", "bucket": "legal-docs"},
    pipeline={
        "embedding": {
            "model": "mixpeek://text_extractor@v1/ibm_granite_embed_311m_multi_r2"
        }
    }
)

Capabilities

200+ language support with 52 enhanced languages
32K token context length via RoPE
768-dimensional embeddings with Matryoshka truncation to 128-dim
Code retrieval across Python, Go, Java, JavaScript, PHP, Ruby, SQL, C, C++
1828 docs/sec throughput on single H100

Use Cases on Mixpeek

Cross-lingual enterprise search across global document repositories

Long-document embedding for legal contracts and research papers without chunking

Multilingual code search across polyglot codebases

Edge-optimized deployment via OpenVINO INT8 quantization

Benchmarks

Dataset	Metric	Score	Source
MTEB Multilingual Retrieval (18 tasks)	nDCG@10	65.2	Model card
MTEB Code Retrieval (12 tasks)	nDCG@10	63.8	Model card
LongEmbed (6 tasks)	nDCG@10	71.7	Model card