README.md

---
library_name: mlx
license: mit
pipeline_tag: feature-extraction
base_model: intfloat/multilingual-e5-large
tags:
- mlx
- embeddings
- sentence-transformers
- xlm-roberta
- multilingual
- quantized
- int6
- q6
- revis
---

# mavis-ai/Multilingual-e5-large-Q6

This repository contains a 6-bit quantized MLX-compatible distribution of `intfloat/multilingual-e5-large`, prepared for use with **R.E.V.I.S.** as a smaller local semantic embedding model.

The model is intended for local text embedding, semantic recall, RAG retrieval, and multilingual semantic search workflows.

## Important Notice

This repository is hosted primarily as a dedicated download source for the R.E.V.I.S. application ecosystem. You are free to download and use this model package for your own local embedding or MLX workflows, subject to the MIT License and the attribution notices included in this repository.

This package is **not** a new embedding model and has **not** been fine-tuned. It is a quantized redistribution of `intfloat/multilingual-e5-large`.

For the original model card, training details, intended usage, and evaluation information, refer to the official upstream model:

- Original model: <https://huggingface.co/intfloat/multilingual-e5-large>
- Base architecture: XLM-RoBERTa large
- Embedding size: 1024

## Quantization

This package stores selected 2D weight tensors using a R.E.V.I.S. MLX-native Q6 format:

- Quantization type: `mlx-native-affine`
- Bits: `6`
- Group size: `64`
- Mode: `affine`
- Stored tensors: packed `.qweight` plus `.scales` and `.biases`
- Expected linear path: `mx.quantized_matmul(..., qweight, scales=scales, biases=biases, group_size=64, bits=6, mode="affine")`
- Expected embedding lookup path: gather the packed rows first, then `mx.dequantize(..., scales=scales, biases=biases, group_size=64, bits=6, mode="affine")`

Typical tensor layout:

```text
encoder.layer.0.attention.self.query.weight.qweight
encoder.layer.0.attention.self.query.weight.scales
encoder.layer.0.attention.self.query.weight.biases
```

The `.qweight` tensors are MLX packed integer tensors, not plain row-major integer arrays. Non-quantized tensors, such as LayerNorm parameters, bias tensors, and other small metadata tensors, are preserved in their original floating-point representation.

This format is optimized for smaller download and storage size than Q8 while preserving very similar keyword-retrieval behavior in R.E.V.I.S. Runtimes should read `quantization.json` for the exact tensor names and quantization parameters before loading the weights.

## Optimized for R.E.V.I.S. (Local Cognitive OS)

We host this model package to serve as the local semantic embedding engine for **R.E.V.I.S.**

**R.E.V.I.S.** is a 100% local Cognitive OS for Multi-Agentic AI. It transforms your Mac devices into a distributed Agentic Swarm via zero-config Wi-Fi clustering, allowing you to run heavy AI workloads—like recursive web research, dynamic RAG generation, and multi-step logic—without killing single-machine performance.

If you are interested in pushing the absolute limits of local AI and open-weight models, check out our project.

- Official Website: <https://mavis-ai.co.jp/revis/>
- Watch the 13-min Raw Demo (Multi-node Dynamic RAG): <https://x.gd/LxaBF>
- Follow our updates on X: <https://x.com/mavis_ai_jp>

## Usage Notes

For retrieval-style tasks, E5 models typically use different text prefixes for queries and passages. R.E.V.I.S. applies its own canonical query and passage formatting internally.

If you use this package outside R.E.V.I.S., refer to the upstream E5 instructions for recommended prompt prefixes and pooling behavior.

## Files

Recommended repository files:

```text
README.md
LICENSE
NOTICE
weights.00.safetensors
config.json
tokenizer.json
tokenizer_config.json
special_tokens_map.json
quantization.json
```

## License

This repository redistributes a quantized package derived from `intfloat/multilingual-e5-large`, which is released under the **MIT License**.

The upstream copyright notice and MIT License text are preserved in `LICENSE`.

Additional attribution and redistribution notes are included in `NOTICE`.

## Attribution

Original model:

```text
intfloat/multilingual-e5-large
https://huggingface.co/intfloat/multilingual-e5-large
```

Original authors / associated paper:

```text
Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, Furu Wei
Multilingual E5 Text Embeddings: A Technical Report
```

R.E.V.I.S. Q6 package:

```text
Prepared and redistributed by MAVIS / R.E.V.I.S.
Quantization: MLX-native affine Q6 package for local MLX embedding runtime
```

## Modification Notice

Compared with the upstream `intfloat/multilingual-e5-large` release, this repository applies the following packaging modification:

```text
Selected 2D weight tensors were quantized to the R.E.V.I.S. MLX-native affine Q6 representation described in `quantization.json`.
```

No fine-tuning, additional training, or architecture-level modification has been applied.