Add MiniMax-H3

docs/source/en/_toctree.yml

@@ -375,6 +375,8 @@
         title: Lumina2Transformer2DModel
       - local: api/models/lumina_nextdit2d
         title: LuminaNextDiT2DModel
+      - local: api/models/minimax_h3_transformer3d
+        title: MiniMaxH3Transformer3DModel
       - local: api/models/mochi_transformer3d
         title: MochiTransformer3DModel
       - local: api/models/motif_video_transformer_3d
@@ -459,6 +461,10 @@
         title: AutoencoderKLLTXVideo
       - local: api/models/autoencoderkl_magvit
         title: AutoencoderKLMagvit
+      - local: api/models/autoencoderkl_minimax_h3
+        title: AutoencoderKLMiniMaxH3
+      - local: api/models/autoencoderkl_minimax_h3_audio
+        title: AutoencoderKLMiniMaxH3Audio
       - local: api/models/autoencoderkl_mochi
         title: AutoencoderKLMochi
       - local: api/models/autoencoderkl_qwenimage
@@ -691,6 +697,8 @@
         title: LTX-2
       - local: api/pipelines/ltx_video
         title: LTXVideo
+      - local: api/pipelines/minimax_h3
+        title: MiniMax-H3
       - local: api/pipelines/mochi
         title: Mochi
       - local: api/pipelines/motif_video
@@ -770,6 +778,8 @@
       title: LCMScheduler
     - local: api/schedulers/lms_discrete
       title: LMSDiscreteScheduler
+    - local: api/schedulers/minimax_h3
+      title: MiniMaxH3Scheduler
     - local: api/schedulers/pndm
       title: PNDMScheduler
     - local: api/schedulers/repaint

docs/source/en/api/models/autoencoderkl_minimax_h3.md

@@ -0,0 +1,38 @@
+<!-- Copyright 2026 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# AutoencoderKLMiniMaxH3
+
+The video variational autoencoder (VAE) model with KL loss used in [MiniMax-H3](https://huggingface.co/MiniMaxAI) by MiniMax. It pairs a causal 3D CNN encoder with a non-causal ViT decoder and compresses 16x spatially and 4x temporally.
+
+Three things set it apart from most autoencoders in the library:
+
+- **Latents are normalized per channel.** There is no `scaling_factor`: a pipeline encodes with `(latent - latents_mean) / latents_std` and decodes with `latent * latents_std + latents_mean`.
+- **The pixel convention is ImageNet-normalized RGB over a `[0, 1]` base range**, not the usual `[-1, 1]`. `encode` expects `(pixel - imagenet_mean) / imagenet_std` and `decode` returns values in that same space, so a pipeline applies `sample * imagenet_std + imagenet_mean` and clamps to `[0, 1]` before postprocessing.
+- **Spatial tiling is on by default.** MiniMax-H3 was released with tiling enabled for both encoding and decoding and the released frames are the blended-tile ones, so turning it off changes the output. Use `enable_tiling` to change the tile geometry and `disable_tiling` to switch it off.
+
+The temporal geometry is fixed by `clip_length` (17 pixel frames per encoder chunk) and `token_drop` (3 trailing latent frames dropped per encode), so `17 * n + 5` pixel frames map to `5 * n + 2` latent frames.
+
+```python
+import torch
+from diffusers import AutoencoderKLMiniMaxH3
+
+vae = AutoencoderKLMiniMaxH3.from_pretrained(
+    "MiniMaxAI/MiniMax-H3", subfolder="vae", dtype=torch.float32
+).to("cuda")
+```
+
+## AutoencoderKLMiniMaxH3
+
+[[autodoc]] AutoencoderKLMiniMaxH3
+    - encode
+    - decode
+    - all

docs/source/en/api/models/autoencoderkl_minimax_h3_audio.md

@@ -0,0 +1,36 @@
+<!-- Copyright 2026 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# AutoencoderKLMiniMaxH3Audio
+
+The audio autoencoder used in [MiniMax-H3](https://huggingface.co/MiniMaxAI) by MiniMax. It is waveform in and waveform out, with no mel front-end and no separate vocoder: a DAC-lineage strided convolutional encoder, a causal-attention projection onto the diffusion latent width, and a BigVGAN decoder.
+
+The encoder hops 800 samples at 32 kHz, i.e. 40 latents per second, so a waveform of `800 * n` samples encodes to `n` latents. Waveforms that are not a whole number of hops are right-padded.
+
+The causal-attention projection goes through the attention dispatcher, so `set_attention_backend` applies to it; its mask is `is_causal=True`, which every backend honours except `_native_npu`, whose kernel takes no causal flag.
+
+The autoencoder is **mono**, and it normalizes latents per channel with `latents_mean` / `latents_std` rather than a scalar `scaling_factor`. MiniMax-H3 carries stereo as two *batch* items, and it always consumes the posterior mean (`latent_dist.mode()`), never a sample.
+
+```python
+import torch
+from diffusers import AutoencoderKLMiniMaxH3Audio
+
+audio_vae = AutoencoderKLMiniMaxH3Audio.from_pretrained(
+    "MiniMaxAI/MiniMax-H3", subfolder="audio_vae", dtype=torch.float32
+).to("cuda")
+```
+
+## AutoencoderKLMiniMaxH3Audio
+
+[[autodoc]] AutoencoderKLMiniMaxH3Audio
+    - encode
+    - decode
+    - all

docs/source/en/api/models/minimax_h3_transformer3d.md

@@ -0,0 +1,41 @@
+<!-- Copyright 2026 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# MiniMaxH3Transformer3DModel
+
+A Diffusion Transformer model for joint video and audio generation, introduced in [MiniMax-H3](https://huggingface.co/MiniMaxAI) by MiniMax.
+
+MiniMax-H3 runs a single stack of blocks over **one packed 1-D sequence** that holds the text conditioning, the conditioning image and video rows, the audio rows and the target video rows at once. Attention is full self-attention over that sequence, so there is no cross-attention and no per-modality block weights. Modality-specific behaviour comes only from the two input patch projections, the per-row modality tag that selects the AdaLN modulation parameters, and the two output heads.
+
+Building the packed layout is the caller's job, which is why the forward signature takes the layout apart from the latents: the `(t, h, w)` position grid, the per-row modality tags, the per-row timestep indices and the three index tensors that address the video, audio and text rows. [`MiniMaxH3Blocks`] and [`MiniMaxH3Ref2VABlocks`] build all of it.
+
+A layout that carries padding rows (tag `-1`) needs a masked attention backend, since those rows are kept in their own attention document by a boolean mask; a padless sequence needs no mask and keeps every backend available.
+
+One repository holds both released checkpoint partitions, so the subfolder is what selects the task: `transformer/` for the text and keyframe tasks, `transformer_ref/` for the omni-reference task.
+
+```python
+import torch
+from diffusers import MiniMaxH3Transformer3DModel
+
+transformer = MiniMaxH3Transformer3DModel.from_pretrained(
+    "MiniMaxAI/MiniMax-H3", subfolder="transformer", dtype=torch.bfloat16
+).to("cuda")
+```
+
+The checkpoint is mixed precision: the two input patch projections, the timestep MLP and the two output heads are float32 while the block stack is bfloat16. `from_pretrained` keeps that layout through `_keep_in_fp32_modules`, so pass `dtype=torch.bfloat16` and let it place the float32 modules rather than casting the model with `.to(torch.bfloat16)` afterwards.
+
+## MiniMaxH3Transformer3DModel
+
+[[autodoc]] MiniMaxH3Transformer3DModel
+
+## MiniMaxH3TransformerOutput
+
+[[autodoc]] models.transformers.transformer_minimax_h3.MiniMaxH3TransformerOutput