| --- |
| license: apache-2.0 |
| library_name: mlx |
| pipeline_tag: feature-extraction |
| base_model: TiGa-RCE/Qwen3-Embedding-8B-MLX-BF16 |
| tags: |
| - mlx |
| - embeddings |
| - feature-extraction |
| - sentence-similarity |
| - quantization |
| - omlx |
| - oq4e |
| - 4-bit |
| --- |
| |
| # Qwen3-Embedding-8B — MLX oQ4e |
|
|
| > **Capacity-first experimental checkpoint:** every Q4-family build in this sweep failed the predeclared minimum aligned-cosine fidelity gate. It is published as an experimental comparator, not as a BF16-equivalent embedding model. |
|
|
| This is the **oQe calibrated mixed-precision affine quantization** checkpoint from a matched local embedding-quantization experiment. It is published with explicit lineage, calibration evidence where applicable, and the bounded evaluation result that accompanied the conversion. |
|
|
| ## Provenance and lineage |
|
|
| - Upstream model: [`Qwen/Qwen3-Embedding-8B`](https://huggingface.co/Qwen/Qwen3-Embedding-8B) |
| - Upstream revision recorded for publication: `1d8ad4ca9b3dd8059ad90a75d4983776a23d44af` |
| - Revision evidence: exact source snapshot retained in the local Hugging Face download metadata |
| - Direct parent: [`TiGa-RCE/Qwen3-Embedding-8B-MLX-BF16`](https://huggingface.co/TiGa-RCE/Qwen3-Embedding-8B-MLX-BF16) |
| - Conversion rule: every quantized checkpoint branches directly from the family MLX BF16 checkpoint; no lossy checkpoint was used to create another. |
| - Quantization: oQe calibrated mixed-precision affine quantization, nominal 4-bit, group size 64 |
| - Local conversion stack: oMLX 0.5.3, mlx-lm 0.31.3, MLX 0.32.0 |
| - Full collection: [MLX Embedding Quantization Matrix](https://huggingface.co/collections/TiGa-RCE/mlx-embedding-quantization-matrix-q-oq-oqe-at-4-6-8-bit-6a68d11afb238d4fe967d70b) |
|
|
| `PROVENANCE.json` contains machine-readable lineage and SHA-256 hashes for the published weight files. The oQe build used a 128-sample, 512-token importance matrix. The retained `oq_imatrix_report.json` records coverage and fallbacks. |
|
|
| ## Bounded local evaluation |
|
|
| | Metric | Result | |
| |---|---:| |
| | Top-1 retrieval | 1.000 | |
| | MRR | 1.000 | |
| | Mean aligned cosine vs BF16 | 0.987689 | |
| | Minimum aligned cosine vs BF16 | 0.983151 | |
| | Score RMSE vs BF16 | 0.012938 | |
| | Queries with rank change | 0 | |
| | Predeclared gate | FAIL | |
|
|
| The evaluation used 24 frozen query/document pairs, the upstream query instruction recipe, last-token pooling, L2 normalization, and direct comparison with vectors from the family BF16 checkpoint. This is an engineering smoke test, not MTEB and not a claim of universal quality. Retrieval success and representation fidelity are reported separately. |
|
|
| ## Runtime scope |
|
|
| This checkpoint targets Apple Silicon through MLX/oMLX. CUDA and PyTorch results are a separate control lane and must not be interpreted as measurements of MLX/Metal kernel performance. |
|
|
| ## License and attribution |
|
|
| Apache-2.0, following the upstream model card. The original model authors retain attribution for the upstream model; this repository contains a local MLX conversion or quantized derivative prepared by TiGa-RCE for reproducibility research. |
|
|
