Granite-Embedding-97M-Multilingual-R2-GGUF

GGUF builds of ibm-granite/granite-embedding-97m-multilingual-r2.

This repository is a mirror, published so that AICoworker (OpenClaw gateway) can ship Granite R2 as its default local embedding model without depending on a third-party repository remaining available. The files are byte-identical to the conversion credited below.

Files

Q8_0 is the recommended default: the retrieval-quality gap to BF16 is negligible while the download is roughly half the size.

Pooling is baked into the file

llama.cpp pools token vectors according to the GGUF's own pooling_type — the caller does not get to choose it at load time. These files declare CLS pooling, which is what Granite R2 was trained for. A mean-pooled re-quant still loads and still returns vectors, but retrieval quality degrades measurably with no other symptom, so keep CLS if you re-quantize.

Usage (node-llama-cpp)

import { getLlama, resolveModelFile } from 'node-llama-cpp';

const file = await resolveModelFile(
  'hf:xnohat/granite-embedding-97m-multilingual-r2-GGUF/granite-embedding-97M-multilingual-r2-Q8_0.gguf',
);
const llama = await getLlama();
const model = await llama.loadModel({ modelPath: file });
const context = await model.createEmbeddingContext();

const { vector } = await context.getEmbeddingFor('Mật khẩu wifi nhà là gì?'); // 384 dims

Granite takes raw text — no query: / passage: prefixes (those are E5 and EmbeddingGemma conventions, and applying them here measurably hurts).

Credits

• Original model: ibm-granite/granite-embedding-97m-multilingual-r2 — IBM Granite, Apache-2.0.
• GGUF conversion: mykor/granite-embedding-97m-multilingual-r2-GGUF.

All credit for the model and for the GGUF conversion belongs to the authors above. This repo only pins availability.