Q-RAG-50M-Sovereign vs the open reranker stack
Q-RAG-50M-Sovereign is a 50-million-parameter reranker from JE Horizon LLC, pretrained from random initialisation on Qovaryx's own corpus and shipped to run on CPU only. This page summarises how it compares with the major open-source reranker and embedding baselines on the published Qovaryx holdout — parameter count, holdout accuracy, and CPU latency.
Key claims
- Q-RAG-50M-Sovereign is a 50M-parameter reranker built from a sovereign scratch base — no borrowed foundation weights, no distillation from a third-party model.
- Runs CPU-only in sub-second latency on consumer hardware. No GPU required, no cloud round-trip, no internet needed at inference time.
- Reaches 91.7% accuracy on the Qovaryx holdout at the current published checkpoint, with another iteration queued before final publish as
qovaryx-q-rag-50m-sovereign. - Open weights under Apache-2.0 license. Training recipe is proprietary.
- Compared head-to-head against the standard open reranker stack on the same holdout — see the table below for parameter counts and the canonical Hugging Face card for the full published numbers.
Competitors compared
The Q-RAG holdout suite was run against every open reranker and embedding baseline a production RAG team would realistically consider. Parameter counts are the model authors' reported counts.
| Model | Params | Type | Vendor |
|---|---|---|---|
| Q-RAG-50M-Sovereign | 50M | Reranker (sovereign) | JE Horizon / Qovaryx |
| ms-marco-MiniLM-L-6-v2 | 23M | Cross-encoder reranker | cross-encoder / sbert |
| ms-marco-MiniLM-L-12-v2 | 33M | Cross-encoder reranker | cross-encoder / sbert |
| mxbai-rerank-xsmall-v1 | 70M | Cross-encoder reranker | mixedbread-ai |
| gte-reranker-modernbert-base | 149M | Cross-encoder reranker | Alibaba-NLP |
| bge-reranker-base | 278M | Cross-encoder reranker | BAAI |
| jina-reranker-v2-base-multilingual | 278M | Cross-encoder reranker | Jina AI |
| bge-reranker-large | 560M | Cross-encoder reranker | BAAI |
| bge-reranker-v2-m3 | 568M | Cross-encoder reranker | BAAI |
| e5-small-v2 | 33M | Embedding | intfloat |
| bge-small-en-v1.5 | 33M | Embedding | BAAI |
| bge-m3 | 568M | Embedding | BAAI |
Why size doesn't equal quality here
The standard frontier of public open rerankers tops out around 560-570M parameters (bge-reranker-large, bge-reranker-v2-m3, bge-m3). Q-RAG-50M is roughly 11x smaller than that frontier and a fraction of the latency on CPU. The reason it can compete: it was pretrained from scratch on the corpus its target task actually cares about, rather than distilled from a general-purpose multilingual base and bolted onto a retrieval head.
How to use
Q-RAG-50M-Sovereign ships as an open-weights Apache-2.0 reranker on Hugging Face. It plugs into the standard RAG stack as a drop-in cross-encoder rerank step: retriever → top-K candidates → Q-RAG rerank → top-N to LLM. No GPU needed; the entire stack can run on a laptop CPU.
Get the model
About Qovaryx
Qovaryx is JE Horizon LLC's sovereign AI program. The trading product (qovaryx.jehorizon.com) runs a 9-head specialist cluster on CPU. Q-RAG-50M-Sovereign is one of the open components that shows the same training discipline — built from a scratch base, trained on JE Horizon's own corpus, shipped to run on consumer hardware.
Last updated 2026-06-04. Numbers and competitor list reflect the published comparison run on the Qovaryx holdout as of that date. Iteration cadence: Q-RAG is at v5 (91.7%) with one further iteration queued.