Org card: list all three quantization families with collection links

README.md

@@ -22,9 +22,14 @@ pinned: false
 
 Community model artifacts from our applied-AI work, validated the way we'd want to find them: pinned base revisions, full quantization manifests, raw benchmark JSON alongside every table, and cards that state exactly what was measured and on which hardware.
 
-Current releases: community quantizations of NVIDIA's Nemotron-3-Embed-8B retrieval model — **FP8** (Ada/Hopper/Blackwell), **NVFP4** (Blackwell-native 4-bit), and **MLX 4-bit** (Apple Silicon).
+Current releases, each grouped in its own collection with the upstream base model:
+
+- **[Unlimited-OCR](https://huggingface.co/collections/shadowrock-io/unlimited-ocr-community-quantizations-6a83eadafbe3369a2f5ca186)** (Baidu, 3.34B MoE OCR): **FP8** at CER 0.0 vs BF16 and the first calibrated **NVFP4**, mean CER 0.0068.
+- **[LocateAnything-3B](https://huggingface.co/collections/shadowrock-io/locateanything-3b-community-quantizations-6a83eadba7a8cf13c1866134)** (NVIDIA, open-vocabulary grounding): **FP8** and **NVFP4**, benchmarked on a COCO val2017 grounding slice.
+- **[Nemotron-3-Embed-8B](https://huggingface.co/collections/shadowrock-io/nemotron-3-embed-8b-community-quantizations-6a7179e5dc8bda86bc069f73)** (NVIDIA, retrieval): **FP8**, **NVFP4**, and **MLX 4-bit** (Apple Silicon), within 0.008 nDCG@10 of NVIDIA's published RTEB scores.
 
 ## Find us
 
 - Web: [shadowrock.io](https://shadowrock.io)
 - LinkedIn: [linkedin.com/company/shadowrock](https://www.linkedin.com/company/shadowrock)
+