EVIE-Preview-4.5B on Apple Silicon: MLX Q4 visual retrieval

[victoruce]

Hi! I released an unofficial MLX adaptation of Tencent EVIE-Preview-4.5B for visual document retrieval on Apple Silicon.

• Text or image queries, 128D multi-vector embeddings and MaxSim scoring.
• 3.03 GB weights: Q4 language backbone, BF16 vision/projection.
• Loads with mlx_vlm.load via a bundled model_file adapter; not yet native MLX-VLM support.
• Two complete public ViDoRe tasks: nDCG@5 97.59 (TabFQuAD) and 68.43 (Economics). This is a small validation, not a full leaderboard submission.

Weights, code, quick start and reproducible evaluation — Apache-2.0, with upstream attribution.

Feedback welcome, especially on a path toward native retrieval support in MLX-VLM. Thanks to the Tencent, MLX and MLX-VLM teams!