diff --git a/docs/source/en/optimization/cache.md b/docs/source/en/optimization/cache.md
index 079f073b73f0..200a601e4dcf 100644
--- a/docs/source/en/optimization/cache.md
+++ b/docs/source/en/optimization/cache.md
@@ -116,7 +116,7 @@ pipe.transformer.enable_cache(config)
 
 [MagCache](https://github.com/Zehong-Ma/MagCache) accelerates inference by skipping transformer blocks based on the magnitude of the residual update. It observes that the magnitude of updates (Output - Input) decays predictably over the diffusion process. By accumulating an "error budget" based on pre-computed magnitude ratios, it dynamically decides when to skip computation and reuse the previous residual.
 
-MagCache relies on **Magnitude Ratios** (`mag_ratios`), which describe this decay curve. These ratios are specific to the model checkpoint and scheduler.
+MagCache relies on **Magnitude Ratios** (`mag_ratios`), which describe this decay curve. These ratios are specific to the model checkpoint and scheduler. The bundled `FLUX_MAG_RATIOS` were measured on FLUX.1; other models, including Flux2 and Flux2 Klein, need their own calibration run.
 
 To use MagCache, you typically follow a two-step process: **Calibration** and **Inference**.
 
diff --git a/src/diffusers/hooks/_helpers.py b/src/diffusers/hooks/_helpers.py
index 9cbe5bc8108f..0f29e97661c4 100644
--- a/src/diffusers/hooks/_helpers.py
+++ b/src/diffusers/hooks/_helpers.py
@@ -175,6 +175,7 @@ def _register_transformer_blocks_metadata():
     from ..models.transformers.transformer_bria import BriaTransformerBlock
     from ..models.transformers.transformer_cogview4 import CogView4TransformerBlock
     from ..models.transformers.transformer_flux import FluxSingleTransformerBlock, FluxTransformerBlock
+    from ..models.transformers.transformer_flux2 import Flux2SingleTransformerBlock, Flux2TransformerBlock
     from ..models.transformers.transformer_hunyuan_video import (
         HunyuanVideoSingleTransformerBlock,
         HunyuanVideoTokenReplaceSingleTransformerBlock,
@@ -246,6 +247,22 @@ def _register_transformer_blocks_metadata():
         ),
     )
 
+    # Flux2
+    TransformerBlockRegistry.register(
+        model_class=Flux2TransformerBlock,
+        metadata=TransformerBlockMetadata(
+            return_hidden_states_index=1,
+            return_encoder_hidden_states_index=0,
+        ),
+    )
+    TransformerBlockRegistry.register(
+        model_class=Flux2SingleTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=None,
+        ),
+    )
+
     # HunyuanVideo
     TransformerBlockRegistry.register(
         model_class=HunyuanVideoTransformerBlock,
diff --git a/src/diffusers/hooks/mag_cache.py b/src/diffusers/hooks/mag_cache.py
index e5f0aaebc01a..4ae1b5764edc 100644
--- a/src/diffusers/hooks/mag_cache.py
+++ b/src/diffusers/hooks/mag_cache.py
@@ -347,7 +347,8 @@ def new_forward(self, module: torch.nn.Module, *args, **kwargs):
                 if diff == 0:
                     residual = out_hidden - in_hidden
                 else:
-                    residual = out_hidden - in_hidden  # Fallback to matching tail
+                    # The tail returned the fused text+image sequence (e.g. Flux2); the image tokens sit at the end.
+                    residual = out_hidden[:, -in_hidden.shape[1] :] - in_hidden
             else:
                 # Fallback for completely mismatched shapes
                 residual = out_hidden
diff --git a/tests/models/transformers/test_models_transformer_flux2.py b/tests/models/transformers/test_models_transformer_flux2.py
index 3263ce68202c..75fe4a1701dd 100644
--- a/tests/models/transformers/test_models_transformer_flux2.py
+++ b/tests/models/transformers/test_models_transformer_flux2.py
@@ -38,6 +38,7 @@
     GGUFTesterMixin,
     LoraHotSwappingForModelTesterMixin,
     LoraTesterMixin,
+    MagCacheTesterMixin,
     MemoryTesterMixin,
     ModelTesterMixin,
     SingleFileTesterMixin,
@@ -713,3 +714,7 @@ def pretrained_model_name_or_path(self):
     @property
     def pretrained_model_kwargs(self):
         return {"subfolder": "transformer"}
+
+
+class TestFlux2TransformerMagCache(Flux2TransformerTesterConfig, MagCacheTesterMixin):
+    """MagCache tests for Flux2 Transformer."""
diff --git a/tests/pipelines/flux2/test_pipeline_flux2.py b/tests/pipelines/flux2/test_pipeline_flux2.py
index ac72d843cd05..9c9bdb2468ef 100644
--- a/tests/pipelines/flux2/test_pipeline_flux2.py
+++ b/tests/pipelines/flux2/test_pipeline_flux2.py
@@ -13,6 +13,7 @@
     BasePipelineTesterConfig,
     LoraMemoryTesterMixin,
     LoraTesterMixin,
+    MagCacheTesterMixin,
     MemoryTesterMixin,
     PipelineTesterMixin,
     check_qkv_fused_layers_exist,
@@ -204,3 +205,7 @@ class TestFlux2PipelineLoRAMemory(Flux2PipelineTesterConfig, LoraMemoryTesterMix
 
     # See `TestFlux2PipelineLoRA`.
     denoiser_target_modules = {"transformer": ["to_qkv_mlp_proj", "to_k"]}
+
+
+class TestFlux2PipelineMagCache(Flux2PipelineTesterConfig, MagCacheTesterMixin):
+    """MagCache tests for the Flux2 pipeline."""
diff --git a/tests/pipelines/flux2/test_pipeline_flux2_klein.py b/tests/pipelines/flux2/test_pipeline_flux2_klein.py
index 0d7139b21e16..84cc08abb53d 100644
--- a/tests/pipelines/flux2/test_pipeline_flux2_klein.py
+++ b/tests/pipelines/flux2/test_pipeline_flux2_klein.py
@@ -23,6 +23,7 @@
 )
 from ..testing_utils import (
     BasePipelineTesterConfig,
+    MagCacheTesterMixin,
     MemoryTesterMixin,
     PipelineTesterMixin,
     check_qkv_fused_layers_exist,
@@ -283,3 +284,7 @@ def test_flux2_klein_neuron_compile_128(self):
         assert image.shape == (1, 128, 128, 3)
         assert not np.isnan(image).any(), "Output contains NaN values"
         assert (image >= 0.0).all() and (image <= 1.0).all(), "Output pixel values outside [0, 1]"
+
+
+class TestFlux2KleinPipelineMagCache(Flux2KleinPipelineTesterConfig, MagCacheTesterMixin):
+    """MagCache tests for the Flux2 Klein pipeline."""