diff --git a/comfy/ldm/kandinsky5/model.py b/comfy/ldm/kandinsky5/model.py
new file mode 100644
index 000000000..064591aa3
--- /dev/null
+++ b/comfy/ldm/kandinsky5/model.py
@@ -0,0 +1,402 @@
+import torch
+from torch import nn
+import math
+
+import comfy.ldm.common_dit
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.math import apply_rope1
+from comfy.ldm.flux.layers import EmbedND
+
+def apply_scale_shift_norm(norm, x, scale, shift):
+    return torch.addcmul(shift, norm(x), scale + 1.0)
+
+def apply_gate_sum(x, out, gate):
+    return torch.addcmul(x, gate, out)
+
+def get_shift_scale_gate(params):
+    shift, scale, gate = torch.chunk(params, 3, dim=-1)
+    return tuple(x.unsqueeze(1) for x in (shift, scale, gate))
+
+def get_freqs(dim, max_period=10000.0):
+    return torch.exp(
+        -math.log(max_period)
+        * torch.arange(start=0, end=dim, dtype=torch.float32)
+        / dim)
+
+class TimeEmbeddings(nn.Module):
+    def __init__(self, model_dim, time_dim, max_period=10000.0, operation_settings=None):
+        super().__init__()
+        assert model_dim % 2 == 0
+        self.model_dim = model_dim
+        self.max_period = max_period
+        self.register_buffer("freqs", get_freqs(model_dim // 2, max_period), persistent=False)
+        self.dtype = operation_settings.get("dtype")
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(model_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=self.dtype)
+        self.activation = nn.SiLU()
+        self.out_layer = operations.Linear(time_dim, time_dim, bias=True, device=operation_settings.get("device"), dtype=self.dtype)
+
+    def forward(self, timestep):
+        args = torch.outer(timestep, self.freqs.to(device=timestep.device))
+        time_embed = torch.cat([torch.cos(args), torch.sin(args)], dim=-1).to(self.dtype)
+        time_embed = self.out_layer(self.activation(self.in_layer(time_embed)))
+        return time_embed
+
+
+class TextEmbeddings(nn.Module):
+    def __init__(self, text_dim, model_dim, operation_settings=None):
+        super().__init__()
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(text_dim, model_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.norm = operations.LayerNorm(model_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, text_embed):
+        text_embed = self.in_layer(text_embed)
+        return self.norm(text_embed).type_as(text_embed)
+
+
+class VisualEmbeddings(nn.Module):
+    def __init__(self, visual_dim, model_dim, patch_size, operation_settings=None):
+        super().__init__()
+        self.patch_size = patch_size
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(visual_dim, model_dim, device=operation_settings.get("device"), dtype=torch.float32)
+
+    def forward(self, x):
+        x = x.movedim(1, -1)  # B C T H W -> B T H W C
+        B, T, H, W, dim = x.shape
+        pt, ph, pw = self.patch_size
+
+        x = x.view(
+            B,
+            T // pt, pt,
+            H // ph, ph,
+            W // pw, pw,
+            dim,
+        ).permute(0, 1, 3, 5, 2, 4, 6, 7).flatten(4, 7)
+
+        return self.in_layer(x.float()).to(x.dtype)
+
+
+class Modulation(nn.Module):
+    def __init__(self, time_dim, model_dim, num_params, operation_settings=None):
+        super().__init__()
+        self.activation = nn.SiLU()
+        operations = operation_settings.get("operations")
+        self.out_layer = operations.Linear(time_dim, num_params * model_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, x):
+        return self.out_layer(self.activation(x))
+
+class SelfAttention(nn.Module):
+    def __init__(self, num_channels, head_dim, operation_settings=None):
+        super().__init__()
+        assert num_channels % head_dim == 0
+        self.num_heads = num_channels // head_dim
+
+        operations = operation_settings.get("operations")
+        self.to_query = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.to_key = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.to_value = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.query_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.key_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        self.out_layer = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def get_qkv(self, x):
+        q = self.to_query(x)
+        k = self.to_key(x)
+        v = self.to_value(x)
+
+        shape = q.shape[:-1]
+        q = q.reshape(*shape, self.num_heads, -1)
+        k = k.reshape(*shape, self.num_heads, -1)
+        v = v.reshape(*shape, self.num_heads, -1)
+
+        return q, k, v
+
+    def forward(self, x, freqs):
+        q, k, v = self.get_qkv(x)
+
+        q = apply_rope1(self.query_norm(q), freqs)
+        k = apply_rope1(self.key_norm(k), freqs)
+
+        out = optimized_attention(
+            q.flatten(-2, -1),
+            k.flatten(-2, -1),
+            v.flatten(-2, -1),
+            heads=self.num_heads)
+        out = self.out_layer(out)
+        return out
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, num_channels, head_dim, operation_settings=None):
+        super().__init__()
+        assert num_channels % head_dim == 0
+        self.num_heads = num_channels // head_dim
+
+        operations = operation_settings.get("operations")
+        self.to_query = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.to_key = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.to_value = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.query_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.key_norm = operations.RMSNorm(head_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        self.out_layer = operations.Linear(num_channels, num_channels, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def get_qkv(self, x, cond):
+        query = self.to_query(x)
+        key = self.to_key(cond)
+        value = self.to_value(cond)
+
+        shape, cond_shape = query.shape[:-1], key.shape[:-1]
+        query = query.reshape(*shape, self.num_heads, -1)
+        key = key.reshape(*cond_shape, self.num_heads, -1)
+        value = value.reshape(*cond_shape, self.num_heads, -1)
+
+        return query, key, value
+
+    def forward(self, x, cond):
+        q, k, v = self.get_qkv(x, cond)
+        q = self.query_norm(q)
+        k = self.key_norm(k)
+
+        out = optimized_attention(
+            q.flatten(-2, -1),
+            k.flatten(-2, -1),
+            v.flatten(-2, -1),
+            heads=self.num_heads)
+
+        out = self.out_layer(out)
+        return out
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, ff_dim, operation_settings=None):
+        super().__init__()
+        operations = operation_settings.get("operations")
+        self.in_layer = operations.Linear(dim, ff_dim, bias=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.activation = nn.GELU()
+        self.out_layer = operations.Linear(ff_dim, dim, bias=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, x):
+        return self.out_layer(self.activation(self.in_layer(x)))
+
+
+class OutLayer(nn.Module):
+    def __init__(self, model_dim, time_dim, visual_dim, patch_size, operation_settings=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.modulation = Modulation(time_dim, model_dim, 2, operation_settings=operation_settings)
+        operations = operation_settings.get("operations")
+        self.norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.out_layer = operations.Linear(model_dim, math.prod(patch_size) * visual_dim, bias=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, visual_embed, time_embed):
+        B, T, H, W, _ = visual_embed.shape
+        shift, scale = torch.chunk(self.modulation(time_embed), 2, dim=-1)
+        scale = scale[:, None, None, None, :]
+        shift = shift[:, None, None, None, :]
+        visual_embed = apply_scale_shift_norm(self.norm, visual_embed, scale, shift)
+        x = self.out_layer(visual_embed)
+
+        out_dim = x.shape[-1] // (self.patch_size[0] * self.patch_size[1] * self.patch_size[2])
+        x = x.view(
+            B, T, H, W,
+            out_dim,
+            self.patch_size[0], self.patch_size[1], self.patch_size[2]
+        )
+        return x.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(2, 3).flatten(3, 4).flatten(4, 5)
+
+class TransformerEncoderBlock(nn.Module):
+    def __init__(self, model_dim, time_dim, ff_dim, head_dim, operation_settings=None):
+        super().__init__()
+        self.text_modulation = Modulation(time_dim, model_dim, 6, operation_settings=operation_settings)
+        operations = operation_settings.get("operations")
+
+        self.self_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.self_attention = SelfAttention(model_dim, head_dim, operation_settings=operation_settings)
+
+        self.feed_forward_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.feed_forward = FeedForward(model_dim, ff_dim, operation_settings=operation_settings)
+
+    def forward(self, x, time_embed, freqs):
+        self_attn_params, ff_params = torch.chunk(self.text_modulation(time_embed), 2, dim=-1)
+        shift, scale, gate = get_shift_scale_gate(self_attn_params)
+        out = apply_scale_shift_norm(self.self_attention_norm, x, scale, shift)
+        out = self.self_attention(out, freqs)
+        x = apply_gate_sum(x, out, gate)
+
+        shift, scale, gate = get_shift_scale_gate(ff_params)
+        out = apply_scale_shift_norm(self.feed_forward_norm, x, scale, shift)
+        out = self.feed_forward(out)
+        x = apply_gate_sum(x, out, gate)
+        return x
+
+
+class TransformerDecoderBlock(nn.Module):
+    def __init__(self, model_dim, time_dim, ff_dim, head_dim, operation_settings=None):
+        super().__init__()
+        self.visual_modulation = Modulation(time_dim, model_dim, 9, operation_settings=operation_settings)
+
+        operations = operation_settings.get("operations")
+        self.self_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.self_attention = SelfAttention(model_dim, head_dim, operation_settings=operation_settings)
+
+        self.cross_attention_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.cross_attention = CrossAttention(model_dim, head_dim, operation_settings=operation_settings)
+
+        self.feed_forward_norm = operations.LayerNorm(model_dim, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.feed_forward = FeedForward(model_dim, ff_dim, operation_settings=operation_settings)
+
+    def forward(self, visual_embed, text_embed, time_embed, freqs):
+        self_attn_params, cross_attn_params, ff_params = torch.chunk(
+            self.visual_modulation(time_embed), 3, dim=-1
+        )
+        # self attention
+        shift, scale, gate = get_shift_scale_gate(self_attn_params)
+        visual_out = apply_scale_shift_norm(self.self_attention_norm, visual_embed, scale, shift)
+        visual_out = self.self_attention(visual_out, freqs)
+        visual_embed = apply_gate_sum(visual_embed, visual_out, gate)
+        # cross attention
+        shift, scale, gate = get_shift_scale_gate(cross_attn_params)
+        visual_out = apply_scale_shift_norm(self.cross_attention_norm, visual_embed, scale, shift)
+        visual_out = self.cross_attention(visual_out, text_embed)
+        visual_embed = apply_gate_sum(visual_embed, visual_out, gate)
+        # feed forward
+        shift, scale, gate = get_shift_scale_gate(ff_params)
+        visual_out = apply_scale_shift_norm(self.feed_forward_norm, visual_embed, scale, shift)
+        visual_out = self.feed_forward(visual_out)
+        visual_embed = apply_gate_sum(visual_embed, visual_out, gate)
+        return visual_embed
+
+
+class Kandinsky5(nn.Module):
+    def __init__(
+        self,
+        in_visual_dim=16, out_visual_dim=16, in_text_dim=3584, in_text_dim2=768, time_dim=512,
+        model_dim=1792, ff_dim=7168, visual_embed_dim=132, patch_size=(1, 2, 2), num_text_blocks=2, num_visual_blocks=32,
+        axes_dims=(16, 24, 24), rope_scale_factor=(1.0, 2.0, 2.0),
+        dtype=None, device=None, operations=None, **kwargs
+    ):
+        super().__init__()
+        head_dim = sum(axes_dims)
+        self.rope_scale_factor = rope_scale_factor
+        self.in_visual_dim = in_visual_dim
+        self.model_dim = model_dim
+        self.patch_size = patch_size
+        self.visual_embed_dim = visual_embed_dim
+        self.dtype = dtype
+        self.device = device
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        self.time_embeddings = TimeEmbeddings(model_dim, time_dim, operation_settings=operation_settings)
+        self.text_embeddings = TextEmbeddings(in_text_dim, model_dim, operation_settings=operation_settings)
+        self.pooled_text_embeddings = TextEmbeddings(in_text_dim2, time_dim, operation_settings=operation_settings)
+        self.visual_embeddings = VisualEmbeddings(visual_embed_dim, model_dim, patch_size, operation_settings=operation_settings)
+
+        self.text_transformer_blocks = nn.ModuleList(
+            [TransformerEncoderBlock(model_dim, time_dim, ff_dim, head_dim, operation_settings=operation_settings) for _ in range(num_text_blocks)]
+        )
+
+        self.visual_transformer_blocks = nn.ModuleList(
+            [TransformerDecoderBlock(model_dim, time_dim, ff_dim, head_dim, operation_settings=operation_settings) for _ in range(num_visual_blocks)]
+        )
+
+        self.out_layer = OutLayer(model_dim, time_dim, out_visual_dim, patch_size, operation_settings=operation_settings)
+
+        self.rope_embedder_3d = EmbedND(dim=head_dim, theta=10000.0, axes_dim=axes_dims)
+        self.rope_embedder_1d = EmbedND(dim=head_dim, theta=10000.0, axes_dim=[head_dim])
+
+    def rope_encode_1d(self, seq_len, seq_start=0, steps=None, device=None, dtype=None, transformer_options={}):
+        steps = seq_len if steps is None else steps
+        seq_ids = torch.linspace(seq_start, seq_start + (seq_len - 1), steps=steps, device=device, dtype=dtype)
+        seq_ids = seq_ids.reshape(-1, 1).unsqueeze(0)  # Shape: (1, steps, 1)
+        freqs = self.rope_embedder_1d(seq_ids).movedim(1, 2)
+        return freqs
+
+    def rope_encode_3d(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, transformer_options={}):
+
+        patch_size = self.patch_size
+        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
+        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
+        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+
+        if steps_t is None:
+            steps_t = t_len
+        if steps_h is None:
+            steps_h = h_len
+        if steps_w is None:
+            steps_w = w_len
+
+        h_start = 0
+        w_start = 0
+        rope_options = transformer_options.get("rope_options", None)
+        if rope_options is not None:
+            t_len = (t_len - 1.0) * rope_options.get("scale_t", 1.0) + 1.0
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+
+            t_start += rope_options.get("shift_t", 0.0)
+            h_start += rope_options.get("shift_y", 0.0)
+            w_start += rope_options.get("shift_x", 0.0)
+        else:
+            # this is experimental as the original code only had two fixed scales for 512p and 1024p
+            if t == 1: # image model
+                self.rope_scale_factor = (1.0, 1.0, 1.0)
+            else:
+                spatial_size = h * w
+                scale_16384 = (1.0, 3.16, 3.16)
+                scale_9216 = (1.0, 2.0, 2.0)
+                if spatial_size <= 6144:
+                    self.rope_scale_factor = scale_9216
+                elif spatial_size >= 14080:
+                    self.rope_scale_factor = scale_16384
+                else:
+                    t = (spatial_size - 14080) / (6144 - 14080)
+                    self.rope_scale_factor = tuple(a + (b - a) * t for a, b in zip(scale_16384, scale_9216))
+
+            t_len = (t_len - 1.0) // self.rope_scale_factor[0] + 1.0
+            h_len = (h_len - 1.0) // self.rope_scale_factor[1] + 1.0
+            w_len = (w_len - 1.0) // self.rope_scale_factor[2] + 1.0
+
+        img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_start, h_start + (h_len - 1), steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_start, w_start + (w_len - 1), steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1)
+        img_ids = img_ids.reshape(1, -1, img_ids.shape[-1])
+
+        freqs = self.rope_embedder_3d(img_ids).movedim(1, 2)
+        return freqs
+
+    def forward_orig(self, x, timestep, context, y, freqs, freqs_text, transformer_options={}, **kwargs):
+        context = self.text_embeddings(context)
+        time_embed = self.time_embeddings(timestep) + self.pooled_text_embeddings(y)
+
+        for block in self.text_transformer_blocks:
+            context = block(context, time_embed, freqs_text)
+
+        visual_embed = self.visual_embeddings(x)
+        visual_shape = visual_embed.shape[:-1]
+
+        for block in self.visual_transformer_blocks:
+            visual_embed = block(visual_embed.flatten(1, -2), context, time_embed, freqs=freqs)
+
+        visual_embed = visual_embed.reshape(*visual_shape, -1)
+        return self.out_layer(visual_embed, time_embed)
+
+    def _forward(self, x, timestep, context, y, transformer_options={}, **kwargs):
+        bs, c, t_len, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+        freqs = self.rope_encode_3d(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options)
+        freqs_text = self.rope_encode_1d(context.shape[1], device=x.device, dtype=x.dtype, transformer_options=transformer_options)
+        return self.forward_orig(x, timestep, context, y, freqs, freqs_text, transformer_options=transformer_options, **kwargs)
+
+
+    def forward(self, x, timestep, context, y, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, y, transformer_options=transformer_options, **kwargs)
diff --git a/comfy/model_base.py b/comfy/model_base.py
index 9b76c285e..9ce40482f 100644
--- a/comfy/model_base.py
+++ b/comfy/model_base.py
@@ -47,6 +47,7 @@ import comfy.ldm.chroma_radiance.model
 import comfy.ldm.ace.model
 import comfy.ldm.omnigen.omnigen2
 import comfy.ldm.qwen_image.model
+import comfy.ldm.kandinsky5.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -1642,3 +1643,48 @@ class HunyuanVideo15_SR_Distilled(HunyuanVideo15):
         out = super().extra_conds(**kwargs)
         out['disable_time_r'] = comfy.conds.CONDConstant(False)
         return out
+
+
+class Kandinsky5(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.kandinsky5.model.Kandinsky5)
+
+    def encode_adm(self, **kwargs):
+        return kwargs["pooled_output"]
+
+    def concat_cond(self, **kwargs):
+        noise = kwargs.get("noise", None)
+
+        image = kwargs.get("concat_latent_image", None)
+        device = kwargs["device"]
+
+        if image is None:
+            image = torch.zeros_like(noise)
+        else:
+            image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            image = self.process_latent_in(image)
+            image = utils.resize_to_batch_size(image, noise.shape[0])
+
+        mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if mask is None:
+            mask = torch.zeros_like(noise)[:, :1]
+        else:
+            mask = 1.0 - mask
+            mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            if mask.shape[-3] < noise.shape[-3]:
+                mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0)
+            mask = utils.resize_to_batch_size(mask, noise.shape[0])
+
+        return torch.cat((image, mask), dim=1)
+
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        return out
diff --git a/comfy/model_detection.py b/comfy/model_detection.py
index 7d0517e61..82b3700e5 100644
--- a/comfy/model_detection.py
+++ b/comfy/model_detection.py
@@ -625,6 +625,24 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
         return dit_config
 
+    if '{}visual_transformer_blocks.0.cross_attention.key_norm.weight'.format(key_prefix) in state_dict_keys: # Kandinsky 5
+        dit_config = {}
+        model_dim = state_dict['{}visual_embeddings.in_layer.bias'.format(key_prefix)].shape[0]
+        dit_config["model_dim"] = model_dim
+        if model_dim in [4096, 2560]: # pro video and lite image
+            dit_config["axes_dims"] = (32, 48, 48)
+            if model_dim == 2560: # lite image
+                dit_config["rope_scale_factors"] = (1.0, 1.0, 1.0)
+        elif model_dim == 1792: # lite video
+            dit_config["axes_dims"] = (16, 24, 24)
+        dit_config["time_dim"] = state_dict['{}time_embeddings.in_layer.bias'.format(key_prefix)].shape[0]
+        dit_config["image_model"] = "kandinsky5"
+        dit_config["ff_dim"] = state_dict['{}visual_transformer_blocks.0.feed_forward.in_layer.weight'.format(key_prefix)].shape[0]
+        dit_config["visual_embed_dim"] = state_dict['{}visual_embeddings.in_layer.weight'.format(key_prefix)].shape[1]
+        dit_config["num_text_blocks"] = count_blocks(state_dict_keys, '{}text_transformer_blocks.'.format(key_prefix) + '{}.')
+        dit_config["num_visual_blocks"] = count_blocks(state_dict_keys, '{}visual_transformer_blocks.'.format(key_prefix) + '{}.')
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
diff --git a/comfy/sd.py b/comfy/sd.py
index f9e5efab5..0bda30fdb 100644
--- a/comfy/sd.py
+++ b/comfy/sd.py
@@ -54,6 +54,7 @@ import comfy.text_encoders.qwen_image
 import comfy.text_encoders.hunyuan_image
 import comfy.text_encoders.z_image
 import comfy.text_encoders.ovis
+import comfy.text_encoders.kandinsky5
 
 import comfy.model_patcher
 import comfy.lora
@@ -958,6 +959,7 @@ class CLIPType(Enum):
     HUNYUAN_IMAGE = 19
     HUNYUAN_VIDEO_15 = 20
     OVIS = 21
+    KANDINSKY5 = 22
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -1208,6 +1210,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif clip_type == CLIPType.HUNYUAN_VIDEO_15:
             clip_target.clip = comfy.text_encoders.hunyuan_image.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer
+        elif clip_type == CLIPType.KANDINSKY5:
+            clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5Tokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer
diff --git a/comfy/supported_models.py b/comfy/supported_models.py
index afd97160b..306e35270 100644
--- a/comfy/supported_models.py
+++ b/comfy/supported_models.py
@@ -21,6 +21,7 @@ import comfy.text_encoders.ace
 import comfy.text_encoders.omnigen2
 import comfy.text_encoders.qwen_image
 import comfy.text_encoders.hunyuan_image
+import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.z_image
 
 from . import supported_models_base
@@ -1474,7 +1475,36 @@ class HunyuanVideo15_SR_Distilled(HunyuanVideo):
         hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideo15Tokenizer, comfy.text_encoders.hunyuan_image.te(**hunyuan_detect))
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2]
 
+class Kandinsky5(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "kandinsky5",
+    }
+
+    sampling_settings = {
+        "shift": 10.0,
+    }
+
+    unet_extra_config = {}
+    latent_format = latent_formats.HunyuanVideo
+
+    memory_usage_factor = 1.8 #TODO
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Kandinsky5(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5Tokenizer, comfy.text_encoders.kandinsky5.te(**hunyuan_detect))
+
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Kandinsky5]
 
 models += [SVD_img2vid]
diff --git a/comfy/text_encoders/kandinsky5.py b/comfy/text_encoders/kandinsky5.py
new file mode 100644
index 000000000..cbdc59d01
--- /dev/null
+++ b/comfy/text_encoders/kandinsky5.py
@@ -0,0 +1,58 @@
+from comfy import sd1_clip
+from .qwen_image import QwenImageTokenizer, QwenImageTEModel, Qwen25_7BVLIModel
+
+class Kandinsky5Tokenizer(QwenImageTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+        # yes the typo "promt" was also in the original template...
+        self.llama_template = "<|im_start|>system\nYou are a promt engineer. Describe the video in detail.\nDescribe how the camera moves or shakes, describe the zoom and view angle, whether it follows the objects.\nDescribe the location of the video, main characters or objects and their action.\nDescribe the dynamism of the video and presented actions.\nName the visual style of the video: whether it is a professional footage, user generated content, some kind of animation, video game or scren content.\nDescribe the visual effects, postprocessing and transitions if they are presented in the video.\nPay attention to the order of key actions shown in the scene.<|im_end|>\n<|im_start|>user\n{}<|im_end|>"
+        self.llama_template_image2video = "<|im_start|>system\nYou are a promt engineer. Your task is to create a highly detailed and effective video description based on a provided input image.\nDescribe how the camera moves or shakes, describe the zoom and view angle, whether it follows the objects.\nDescribe main characters actions.\nDescribe the dynamism of the video and presented actions.\nName the visual style of the video: whether it is a professional footage, user generated content, some kind of animation, video game or scren content.\nDescribe the visual effects, postprocessing and transitions if they are presented in the video.\nPay attention to the order of key actions shown in the scene.<|im_end|>\n<|im_start|>user\n{}<|im_end|>"
+        self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
+        out = super().tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs)
+
+        return out
+
+
+class Kandinsky5TEModel(QwenImageTEModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super(QwenImageTEModel, self).__init__(device=device, dtype=dtype, name="qwen25_7b", clip_model=Qwen25_7BVLIModel, model_options=model_options)
+        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options)
+
+    def encode_token_weights(self, token_weight_pairs):
+        #tok_pairs = token_weight_pairs["qwen25_7b"][0]
+        token_weight_pairs_l = token_weight_pairs["l"]
+        template_end = -1
+
+        cond, p, extra = super().encode_token_weights(token_weight_pairs, template_end=template_end)
+        l_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs_l)
+        
+        return cond, l_pooled, extra
+
+    def set_clip_options(self, options):
+        super().set_clip_options(options)
+        self.clip_l.set_clip_options(options)
+
+    def reset_clip_options(self):
+        super().reset_clip_options()
+        self.clip_l.reset_clip_options()
+
+    def load_sd(self, sd):
+        if "text_model.encoder.layers.1.mlp.fc1.weight" in sd:
+            return self.clip_l.load_sd(sd)
+        else:
+            return super().load_sd(sd)
+
+def te(dtype_llama=None, llama_scaled_fp8=None):
+    class Kandinsky5TEModel_(Kandinsky5TEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["qwen_scaled_fp8"] = llama_scaled_fp8
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return Kandinsky5TEModel_
diff --git a/nodes.py b/nodes.py
index 356aa63df..61f06eb92 100644
--- a/nodes.py
+++ b/nodes.py
@@ -970,7 +970,7 @@ class DualCLIPLoader:
     def INPUT_TYPES(s):
         return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ),
                               "clip_name2": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15"], ),
+                              "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),