A Q8_0 GGUF conversion of lightonai/DenseOn, a 149M-parameter ModernBERT dense retrieval model. DenseOn emits one 768-dimensional, L2-normalized vector using CLS pooling and was trained with asymmetric query/document prefixes.

Files

File	Purpose	SHA-256
`DenseOn-Q8_0.gguf`	Q8_0 ModernBERT encoder with native CLS pooling metadata	`506d5bab02a4adf7c00a71fef2bb58bfa54f769e15670886e212af7c16b4ccff`

litembeddings usage

sql

.load ./litembeddings

SELECT lembed_model(
  '/path/to/DenseOn-Q8_0.gguf',
  json_object('ctx_size', 512, 'batch_size', 512)
);

-- Prefixes are required by DenseOn's training contract.
SELECT lembed('query: best database indexing strategy');
SELECT lembed('document: covering indexes can avoid table lookups');

Use query: for queries and document: for corpus passages. Omitting or swapping these prefixes can reduce retrieval quality. The source model's maximum sequence length is 512 tokens.

Conversion provenance

Source: lightonai/DenseOn
Source revision: cb9947ebccb33862d24e3c7ca2edb25e51acd887
Converted: 2026-07-14
llama.cpp revision: 6eddde06a4f25d55d538b5d15628dcc2b6882147
Quantization: Q8_0
Pooling: native CLS pooling from sentence-transformers metadata

An end-to-end parity check against the source FP32 Transformers pipeline produced cosine similarity 0.999389 for the validation query. Small differences are expected from Q8_0 quantization.

License and attribution

The source model is released under Apache 2.0. See the DenseOn model card for training details, evaluation results, intended use, limitations, and citation information. This repository is an independent conversion and is not affiliated with LightOn.

embedme/lightonai-denseon-Q8_0

DenseOn Q8_0 for litembeddings

Files

litembeddings usage

Conversion provenance

License and attribution

embedme/
lightonai-denseon-Q8_0