Adds ColTurk-VDR-Qwen3VL-4B-v1.0 — a ColBERT-style late-interaction visual document retriever built on Qwen/Qwen3-VL-4B-Instruct (LoRA r=32, published as merged full weights; colpali-engine ColQwen3 architecture, Apache-2.0, open weights, public training code and data).

Since the model is a colpali-engine-native checkpoint (not a trust_remote_code repo), this PR also adds a small ColQwen3EngineWrapper that mirrors the existing ColQwen2_5Wrapper pattern (delegates to colpali_engine.models.ColQwen3 / ColQwen3Processor via ColPaliEngineWrapper).

ViDoRe V3 results (8 public retrieval tasks, full corpus, all queries, MaxSim) — mean NDCG@10 = 0.5584:

Eval code & raw result JSONs (with seeded bootstrap 95% CIs): https://github.com/Verm1lion/ColTurk-VDR — results PR: embeddings-benchmark/results#565 · private-split request: #4797.

• I have filled out the ModelMeta object to the extent possible
• I have ensured that my model can be loaded using
  • mteb.get_model(model_name, revision) and
  • mteb.get_model_meta(model_name, revision)
• I have tested the implementation works on a representative set of tasks.
• The model is public, i.e., is available either as an API or the weights are publicly available to download
• I reproduced results from the original paper (if applicable) on at least one benchmark, and I am including the results in the PR description.

Notes for reviewers:

• The wrapper's load path (ColQwen3.from_pretrained + ColQwen3Processor.from_pretrained on the published repo) was smoke-verified end-to-end on the exact revision pinned in the ModelMeta.
• The results above were produced with the published self-contained harness (identical loading + MaxSim over the full corpora, all queries). Happy to re-run anything via mteb.evaluate on request.
• Training data (the ColPali train set) vs. ViDoRe V3 contamination was checked empirically (perceptual-hash scan over train images x V3 corpora: 0 exact duplicates; report in the linked repo).