webAI

A New Model for Intelligence. The next era of AI won't come from a single system in the cloud, but from a civilization of models owned by the people closest to the problem.

Most retrieval systems break on the documents that matter most: charts, tables, scanned pages, dense layouts. We built one that doesn't. Two of the top three spots on the ViDoRe V3 leaderboard. It's open source. Here's how we trained webAI-ColVec1. 🧵

9/ That approach earned two of the top three spots on ViDoRe V3: heterogeneous data, targeted rebalancing, efficient adaptation, and an objective built for ranking quality. We're open sourcing both models so anyone can build with them. Blog: webai.com/blog/webai-col… Try both models on Hugging Face: ColVec1-9B: huggingface.co/webAI-Official… ColVec1-4B: huggingface.co/webAI-Official…

8/ The loss function is proprietary. But the principle: force the model to create cleaner separation between the correct page and every competing page in the batch. Better embedding geometry. Cleaner ranking. Especially important when pages look semantically similar.

7/ Why batch size matters here: we use in-batch negatives. For every query, its matched page is the positive. Every other page in the batch is a negative. With an effective batch size of 512, each query is learning against 511 competing document pages per step. Larger batch = more negatives = stronger contrastive signal.

6/ For adaptation, we used LoRA (rank 32, alpha 32, dropout 0.1) on the Qwen 3.5 backbone plus a retrieval projection layer on top. This gives us a lightweight way to specialize the model for retrieval without full fine-tuning. Trained on 8 A100s with an effective batch size of 512.

5/ We didn't just merge the data, we rebalanced it. ColPali train set duplicated 2x. VDR-multilingual filtered to English and French, then duplicated 5x. Deliberate weighting to increase the influence of the most valuable supervision. The mix matters as much as the volume.

4/ Training data was deliberately heterogeneous. We combined three sources: – Internally collected domain corpora (AI, energy, government, healthcare) – Public document datasets adapted into retrieval format (PubTables, ArxivQA, DocVQA, InfoVQA, TAT-DQA) – Public retrieval datasets (ColPali, VDR-multilingual, VisRAG)

3/ Built on Qwen 3.5 vision-language backbones. We trained multiple variants across 4B and 9B, with embedding sizes of 128, 640, and 2560. Smaller embeddings for speed-sensitive deployments. Larger embeddings when retrieval quality is the priority. The right tradeoff depends on the use case, so we built for both.

2/ webAI-ColVec1 takes a different path. Instead of converting pages into text through a lossy OCR layer, it consumes the raw page image directly and learns retrieval features from the document as it actually appears. The model sees the layout, the tables, the charts, the structure. The same way you do when you look at a page.

1/ Most retrieval systems convert PDFs to text first. OCR, then search. That adds preprocessing, latency, and error propagation. When OCR quality drops, retrieval breaks. And when retrieval breaks, everything downstream breaks too. Your RAG pipeline is only as good as what it retrieves.

We're open-sourcing webAI-ColVec1. #1 on ViDoRe V3. Two of the top three spots. For a long time, multimodal RAG looked like a scaling problem. It's not. Frontier-level retrieval. No OCR. No preprocessing. Built for real documents.

Read the full blog: webai.com/blog/webai-col… Try both models on @huggingface: ColVec1-9B – huggingface.co/webAI-Official… ColVec1-4B – huggingface.co/webAI-Official…

@thewebAI now has 2 of the top 3 RAG models on ViDoRe. And both can run on an iPhone or your laptop. Our smallest model (4B) scores within 0.1 nDCG of @nvidia's nemotron-colembed-vl-8b-v2. That’s within margin of error. Same retrieval quality. Half the model. For a long time, multimodal RAG looked like a scaling problem: Bigger models More GPUs More data center spend That assumption just broke. We’re now seeing: - frontier-level retrieval (64.48 - 63.49 nDCG) - real multimodal understanding (documents, not just text) - running locally This isn’t just a leaderboard result. It means: - your data doesn’t need to leave your device - latency drops to device speed - marginal cost goes to zero - retrieval becomes private by default The best RAG models now fit in your pocket. And we’re making them available. So anyone can build, run, and own their retrieval stack locally. Available: lnkd.in/ejxvZyFe lnkd.in/ejfgVBYP

More open source announcements this afternoon with @thewebAI . This one is very special. #ai #opensource

@berryxia You nailed it. More coming soon 🤝

兄弟们，webAI 首款开源重磅来了！ YOLO26-MLX 正式发布：纯原生跑在 Apple Silicon 上，彻底告别 PyTorch 和外接 GPU！ ✅ 推理速度最高提升 2.6 倍 ✅ 训练速度最高提升 1.7 倍 ✅ 精度仅比官方差 0.2% 纯 MLX 实现，Mac 用户直接本地跑最强 YOLO26，太丝滑了🤯 这是 webAI 开源第一弹，后面还有更多！ 地址见评论区👇

@jtdavies Thanks for sharing John! Very curious to see what results you get on that M5 Max 👀

Wow, amazing! I've been using YOLO for years, mainly on Raspberry Pis to track people, foxes, cats, bikes and cars around my house. It was already blisteringly fast but MLX integration is fantastic. Looking forward to trying it on my M5 Max.

@Dhruv_Syam Yes! Apple Silicon is removing those barriers so anyone can build, definitely let us know how it goes when you get a chance to run it!

@thewebAI got started with cv through ml5.js because my computer couldn't run anything else. really cool to see native apple silicon inference like this!

Our first open-source release. YOLO26-MLX, native YOLO26 on Apple Silicon. No PyTorch. No external GPU. Up to 2.6x faster inference. Up to 1.7x faster training. Accuracy within 0.2% of official results. It won't be the last.