GGUF builds of IdealisticSolutions/islamqa-retriever-quran, an Arabic embedding model (bi-encoder) for Qur'anic passage retrieval. See the main repo's model card for metrics, the Qur'an/Hadith trade-off, and usage policy.

This is an embedding model, not a chat model. It converts Arabic text into 768-dimensional vectors for semantic search. It cannot generate text or hold a conversation — use it as the embedding/retrieval component of a RAG system (LM Studio: load it as an embedding model; Ollama/llama.cpp: the /api/embed / llama-server --embedding endpoints).

File	Precision	Size	Note
`islamqa-retriever-quran-f16.gguf`	F16	287 MB	recommended — max cosine deviation vs original: 0.0003
`islamqa-retriever-quran-q8_0.gguf`	Q8_0	154 MB	smaller; max cosine deviation: 0.005

Architecture: modern-bert (requires a llama.cpp runtime from Dec 2025 or newer)
Pooling: mean (encoded in GGUF metadata) · dim 768 · max input 512 tokens (over-length input is rejected, not truncated — chunk long passages)
No prompt prefixes/instructions — encode plain normalized Arabic text
Parity vs the original sentence-transformers checkpoint: cosine ≥ 0.9999 per sentence on real Qur'an/Hadith text (F16, verified 2026-08-20)

License: Apache-2.0, derived from NAMAA-Space/AraModernBert-Base-STS.

Troubleshooting

LM Studio / llama.cpp chat error "the current context does not logits computation" (HTTP 500): you sent a chat message to an embedding model. This model cannot generate text anywhere, by design. Use it as the embedding model for RAG / document search, or via POST /v1/embeddings.
"unknown architecture: modern-bert": your runtime is too old — modern-bert entered llama.cpp in December 2025. Update LM Studio / your llama.cpp build.
Input over 512 tokens is rejected (not truncated) — chunk long passages before embedding.

IdealisticSolutions/islamqa-retriever-quran-GGUF

islamqa-retriever-quran — GGUF

Troubleshooting

IdealisticSolutions/
islamqa-retriever-quran-GGUF