---
title: WeMM-Embedding-2B-Apple-Q8-G64
canonical_url: "https://www.modelscope.cn/models/xiaotianfotos/WeMM-Embedding-2B-Apple-Q8-G64"
md_url: "https://www.modelscope.cn/models/xiaotianfotos/WeMM-Embedding-2B-Apple-Q8-G64.md"
repository: xiaotianfotos/WeMM-Embedding-2B-Apple-Q8-G64
chinese_name: "Indexed 专用 WeMM 2B Apple Q8/G64"
last_updated: 2026-09-07
license: apache-2.0
pipeline_tag: feature-extraction
tasks:
  - feature-extraction
base_model:
  - tencent/WeMM-Embedding-2B
base_model_relation: finetune
parameters: 529.5M
tensor_type:
  - U32
  - BF16
library_name:
  - mlx
  - safetensors
downloads: 83
stars: 0
tags:
  - indexed
  - mlx
  - coreml
  - multimodal-embedding
  - quantized
---

# WeMM-Embedding-2B-Apple-Q8-G64

> WeMM-Embedding-2B-Apple-Q8-G64 - xiaotianfotos 在 ModelScope 开源的模型。Indexed 开源多模态检索软件专用模型包，基于 Tencent WeMM-Embedding-2B，包含 MLX Q8/G64 语言权重与 Core ML 视觉编码器。使用 Indexed 原生 Swift 后端，无需 Python。项目：https://github.com/xiaotianfotos/indexed

xiaotianfotos/WeMM-Embedding-2B-Apple-Q8-G64 是 ModelScope 魔搭社区上的 529.5M 参数feature-extraction模型，采用 apache-2.0 许可，基于 tencent/WeMM-Embedding-2B 构建。

- **Repository**: xiaotianfotos/WeMM-Embedding-2B-Apple-Q8-G64
- **License**: apache-2.0
- **Tasks**: feature-extraction
- **Parameters**: 529.5M
- **Base model**: tencent/WeMM-Embedding-2B
- **Tags**: indexed, mlx, coreml, multimodal-embedding, quantized
- **Downloads**: 83
- **Stars**: 0
- **Last updated**: 2026-09-07

Source: https://www.modelscope.cn/models/xiaotianfotos/WeMM-Embedding-2B-Apple-Q8-G64

---

# WeMM-Embedding-2B Apple Q8/G64 for Indexed

This is a dedicated model package for [Indexed](https://github.com/xiaotianfotos/indexed), an open-source,
Apache-2.0 local multimodal search application. Download the application source
and use its native Swift Apple embedding backend to load this package.

This conversion derives from [Tencent WeMM-Embedding-2B](https://huggingface.co/tencent/WeMM-Embedding-2B).
It contains MLX affine Q8 weights (group size 64), a compiled Core ML vision
encoder for 448-pixel inputs, the tokenizer, preprocessing configuration,
embedding templates, and the Indexed package manifest. Language weights omit
the generation head and duplicate vision weights. The config's obsolete
Python auto_map has been removed; model tensors and package identity are preserved.

## Runtime and compatibility

The whole repository is an Indexed-specific hybrid package. It is not a
standalone Transformers, mlx-lm, or generic MLX model download. Its supported
product runtime is Swift and requires no Python, pip, or model conversion at
installation. Core ML and MLX execution scheduling belongs to Indexed code.

A/B/C/D share this package. B is the default: Core ML vision with CPU/Neural
Engine routing and MLX GPU language. C/D are experimental and use private ANE
APIs in Indexed; model files alone do not provide those modes. See the
application's diagnostics for actual routing and fallback.

Real-model validation currently covers base Apple M4 on macOS 26. Other Apple
chips and OS versions require separate validation, especially for the compiled
Core ML asset and private ANE paths. No cross-device performance guarantee is made.

## Integrity and installation

Preserve the directory layout and all tokenizer/template files. Install the
local package through Indexed's Apple model setup. Versioned automatic download
integration is not yet provided by the application.

To verify every distributed file from an Indexed source checkout:

`node scripts/prepare-apple-model-package.mjs --verify /path/to/model-package`

The checksums cover configuration, tokenizer, templates, notices and both model
components. Use checksums from a trusted, pinned repository revision; a checksum
file by itself is not a signature. The current native installer separately checks
language and vision hashes from manifest.json; the complete distribution check
above is an additional verification step.

Package fingerprint: `2d481d864704eaabf57a57e77544a6439dd8a524752f74ad5b8543b7f7fe3ecb`.
The manifest defines the embedding-space identity and supported output dimensions.
Do not mix vectors from another package or embedding space in an existing index.

## License and attribution

Retain the complete LICENSE and NOTICE files. Tencent's original model and
third-party components retain their respective license terms. This is a
community conversion for Indexed, not an official Tencent or Apple release.