diff --git a/docs/source/images/nanotabicl_dark.svg b/docs/source/images/nanotabicl_dark.svg
new file mode 100644
index 0000000000..c673300e30
--- /dev/null
+++ b/docs/source/images/nanotabicl_dark.svg
@@ -0,0 +1,4 @@
+
+
+
diff --git a/docs/source/images/nanotabicl_light.svg b/docs/source/images/nanotabicl_light.svg
new file mode 100644
index 0000000000..e90d5d8a11
--- /dev/null
+++ b/docs/source/images/nanotabicl_light.svg
@@ -0,0 +1,4 @@
+
+
+
diff --git a/docs/source/nanotabicl.md b/docs/source/nanotabicl.md
index 99535ceb2c..12e88c661b 100644
--- a/docs/source/nanotabicl.md
+++ b/docs/source/nanotabicl.md
@@ -4,10 +4,20 @@ The [`sdm.nn`](api/nn) package provides reusable, GPU-accelerated building block
In this guide, we show how to build a minimal [TabICLv2](https://arxiv.org/abs/2602.11139)-inspired Tabular Foundation Model (TFM) with it.
Like other [in-context learning (ICL)](icl) models, it receives labeled context rows together with unlabeled query rows, and predicts the query targets in a single forward pass.
-TabICL pioneered the **"Compress-then-ICL"** approach which has since been adopted by other TFMs such as [TabPFN-3](https://arxiv.org/abs/2605.13986), [TabFM](https://arxiv.org/abs/2609.37959), and [KumoTabular](https://huggingface.co/blog/nvidia/kumo-tabular).
+[TabICL](https://arxiv.org/abs/2502.05564) pioneered the **"Compress-then-ICL"** approach which has since been adopted by other TFMs such as [TabPFN-3](https://arxiv.org/abs/2605.13986), [TabFM](https://arxiv.org/abs/2609.37959), and [KumoTabular](https://huggingface.co/blog/nvidia/kumo-tabular).
Its main advantage is its efficient ICL stage that operates on compressed row representations rather than individual cells.
It consists of three stages:
+```{figure} images/nanotabicl_light.svg
+:figclass: light-only
+:width: 100%
+```
+
+```{figure} images/nanotabicl_dark.svg
+:figclass: dark-only
+:width: 100%
+```
+
1. A **row-wise compression block** that maps each table row to a fixed-size embedding.
2. An **in-context learning block** that models interactions among labeled context rows and uses that context to update query representations.
3. A final **prediction head**.