AX-Qwen3-Embedding-8B-MLX-AXQ-6bit

Research development artifact. Base affine6/group32 packing with higher-precision protection, not MXFP6 or uniformly six-bit storage. No quality certificate is issued.

Property	Value
Source	Qwen/Qwen3-Embedding-8B
Immutable source revision	1d8ad4ca9b3dd8059ad90a75d4983776a23d44af
Complete storage	6,781,332,826 bytes
Complete measured BPW	7.169095
Manifest budget class	7p2bpw
MTP	No MTP payload is advertised.

The precision distribution and exact per-module settings are recorded in axquant_manifest.json, axquant_plan.json, and config.json. Complete BPW includes protected tensors and storage overhead. File size does not establish peak runtime memory requirements. Converted from the pinned original source with existing precision protection floors.

Observed runtime evidence

The exact runtime payload passed the following scoped check on Metal. The receipt binds its runtime input files. This is a functional smoke, not a quality benchmark, MTEB evaluation, coding evaluation, long-context test or MTP acceleration result.

{
  "task": "embedding",
  "runtime": "mlx-vlm",
  "version": "0.7.9",
  "mlx_version": "0.32.4",
  "cases": [
    {
      "finite": true,
      "mars_ranks_above_venus": true,
      "min_cosine_vs_source": 0.99609375,
      "passed": true,
      "shape": [
        3,
        4096
      ],
      "unit_norm": true
    }
  ]
}

Runtime	Evidence for this artifact
MLX-VLM 0.7.9	Passed the task-specific smoke above
MLX-LM	Not separately tested
AX Engine	Not tested; architecture metadata is not execution evidence
omlx	unsupported (static only)
mtplx	unsupported (static only)

Unexercised modalities and MTP behavior remain unverified. Peer runtime-specific exports require distinct variants and their own testing. See the manifest-bound static report.

Use and provenance

python -m pip install -U huggingface_hub mlx-vlm
hf download AutomatosX/AX-Qwen3-Embedding-8B-MLX-AXQ-6bit --local-dir ./AX-Qwen3-Embedding-8B-MLX-AXQ-6bit

Pin the immutable Hub revision. Use mlx_vlm.embedding_loader.load_embedding_model for embedding packs, preserving the source pooling and task prefixes. Other packs require the architecture-aware MLX-VLM loader and source processor/template. The source license and notices are retained where supplied. The upstream repository and its license govern downstream use; no additional commercial grant or upstream benchmark/certification transfer is implied.