bge-m3, F16 GGUF, for ODXBrain

This is BAAI/bge-m3 converted to GGUF at F16 with llama.cpp. It isn't requantised, fine-tuned or edited in any way. It's the exact model file that TerraKernel's ODXBrain retrieval engine was built and measured with, published so you can download it instead of converting it yourself.

bge-m3 was made by BAAI. They haven't reviewed or endorsed this conversion.

What ODXBrain is

ODXBrain is a retrieval engine for AI agents that work with Odoo ERP. You ask it in plain language and get back ranked identifiers, never generated text:

"which invoice is not paid yet"  β†’  account.move, account.payment, account.payment.term

It answers two questions:

  • Which Odoo model holds this data? One question in, three ranked model names out, filtered to what your Odoo version and edition actually have. An agent can then query the right model instead of guessing among hundreds.
  • How do we do this here? A task in, the matching standard operating procedure out, from the SOPs your team has approved. They're managed with odxdocs.

It runs on your own servers with no GPU and no internet connection, and answers a query in about 40 ms on two CPU cores. It's air-gapped by design: ODXBrain is an HTTP server with no HTTP client in it, so nothing in it can call out.

The Odoo knowledge it searches is published separately as an open corpus, terrakernel/odxbrain-corpus (CC0).

Where this model fits

ODXBrain turns each question into a vector with this model, then finds the nearest Odoo model descriptions in a corpus pack built with the same model file. The question and the corpus have to be embedded by identical weights, or their vectors don't compare. That's why ODXBrain names one exact file, and why this repository exists.

We chose bge-m3 by measuring candidates against ODXBrain's own test questions. It was picked for three reasons:

  • It's multilingual. Users ask in their own language. Accuracy is measured in English, German, Greek, Spanish, French, Polish, Russian and Chinese.
  • It's MIT-licensed. Enterprise customers review the model licence, and a non-commercial licence would rule a model out.
  • It runs well on a CPU. That matters for sites with no GPU.

Until now, every customer had to download BAAI's PyTorch weights and convert them with exactly the right converter version. This repository publishes the result of that conversion directly.

The file

bge-m3-f16.gguf   1,157,671,008 bytes
SHA-256           1257ef7bb76105793f89307b02d6b9c29a444d629f9ff0ad56ee023222fa105d

Check the hash before you use it. Other bge-m3 GGUFs on Hugging Face aren't this file. Most are 192 bytes larger because a different converter version wrote different metadata. ODXBrain's corpus packs are built and tested against this exact file. A different file still loads and still returns vectors, so nothing errors, but nobody has measured how accurate it is.

This file never changes. If it's ever reconverted, the new file gets a new name. Pin the repository revision your ODXBrain documentation gives you.

Download

For an air-gapped site, download on a machine that has internet access, check it, then carry it across:

hf download terrakernel/odxbrain-bge-m3-gguf bge-m3-f16.gguf \
  --revision 582d13afe777ea05e979a07890d6b3e6ae8dd975 --local-dir .
sha256sum bge-m3-f16.gguf
# 1257ef7bb76105793f89307b02d6b9c29a444d629f9ff0ad56ee023222fa105d  bge-m3-f16.gguf

ODXBrain never downloads anything. It reads the file from a local path, for example a read-only volume in Docker:

--model /models/bge-m3-f16.gguf --model-id bge-m3

The full set-up is on the ODXBrain Docker Hub page.

What's in it

Architecture XLM-RoBERTa (GGUF bert), 24 layers, 567M parameters
Embedding size 1024
Context 8192 tokens
Weights F16
Tokenizer SentencePiece, 250,002 tokens, embedded in the file

Only the dense embedding is here. bge-m3's sparse and multi-vector (ColBERT) heads are separate files in BAAI's repository, and they weren't converted.

There's no pooling setting in the file. bge-m3's dense embedding is the CLS token, L2-normalised. ODXBrain does that by default. With llama.cpp directly, set CLS pooling yourself (--pooling cls).

How it was made

You can rebuild it byte for byte, which is also how to check that this file is what we say it is. You need Python 3.10 or newer, and it downloads about 2.3 GB.

git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
git checkout 542e9202d7d14863562e1e908be0bf32aeec2c2a
python3 -m venv .venv && . .venv/bin/activate
pip install -r requirements/requirements-convert_hf_to_gguf.txt huggingface_hub

hf download BAAI/bge-m3 --revision 5617a9f61b028005a4858fdac845db406aefb181 \
  config.json tokenizer.json tokenizer_config.json special_tokens_map.json \
  sentencepiece.bpe.model pytorch_model.bin --local-dir bge-m3
python convert_hf_to_gguf.py bge-m3 --outtype f16 --outfile bge-m3-f16.gguf
sha256sum bge-m3-f16.gguf

The result is the SHA-256 above. We converted it on macOS in September 2026, then again on Linux, and got the same bytes both times.

Licence

MIT, the same as bge-m3 and everything it's built on. Attribution and the licence text are in NOTICE.

The MIT licence covers this model file only. The ODXBrain engine has its own licence, which this repository doesn't grant.

Citation

If you use bge-m3, cite BAAI's paper:

@misc{bge-m3,
  title={BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation},
  author={Jianlv Chen and Shitao Xiao and Peitian Zhang and Kun Luo and Defu Lian and Zheng Liu},
  year={2024},
  eprint={2402.03216},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}
Downloads last month
64
GGUF
Model size
0.6B params
Architecture
bert
Hardware compatibility
Log In to add your hardware

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for terrakernel/odxbrain-bge-m3-gguf

Base model

BAAI/bge-m3
Quantized
(315)
this model

Paper for terrakernel/odxbrain-bge-m3-gguf