ColQwen3.5-v3 release tops ViDoRe leaderboard

Strong result, but the real story is efficiency, not raw accuracy. This looks like a benchmark win that matters because it makes a high-performing retrieval model meaningfully cheaper to deploy.

• –The headline number is 75.67 mean on MTEB ViDoRe V3, with a smaller 4.5B model and 320-dim embeddings, which is a very practical tradeoff for document retrieval stacks.
• –V3 only nudges past V2 on English @5, so the author’s own read is persuasive: the big breakthrough was the earlier jump in V2, not a dramatic V3 leap.
• –The public optimization trail is unusually transparent for a model release, which makes the leaderboard claim feel more credible than the usual one-off “SOTA” post.
• –Official support in `colpali-engine` and `vLLM` matters here, because retrieval models are only useful if they are easy to serve.
• –This is a strong fit for teams building vision-RAG or document search systems, where memory footprint and embedding width often matter as much as benchmark rank.