Register Qwen 3.5 MoE Model Presets to Hub (keras-team#2798)

Author: laxmareddyp

keras_hub/src/models/qwen3_5_moe/qwen3_5_moe_presets.py

@@ -1,3 +1,51 @@
 """Qwen3.5 MoE model preset configurations."""
 
-backbone_presets = {}
+backbone_presets = {
+    # Qwen3.5 MoE Model Presets
+    "qwen3_5_moe_35b_a3b_base": {
+        "metadata": {
+            "description": (
+                "35 billion total parameter Qwen3.5 MoE base model "
+                "with ~3 billion active parameters per token. Features "
+                "a 3:1 hybrid attention stack (GatedDeltaNet linear "
+                "attention and full attention) with sparse "
+                "Mixture-of-Experts feedforward for highly efficient "
+                "inference. Supports text and multimodal inputs."
+            ),
+            "params": 35107181936,
+            "path": "qwen3_5_moe",
+        },
+        "kaggle_handle": "kaggle://keras/qwen3-5-moe/keras/qwen3_5_moe_35b_a3b_base/1",
+    },
+    "qwen3_5_moe_35b_a3b": {
+        "metadata": {
+            "description": (
+                "35 billion total parameter Qwen3.5 MoE instruction-tuned "
+                "model with ~3 billion active parameters per token. "
+                "Features a 3:1 hybrid attention stack (GatedDeltaNet "
+                "linear attention and full attention) with sparse "
+                "Mixture-of-Experts feedforward. Optimized for chat, "
+                "reasoning, coding, and multimodal tasks."
+            ),
+            "params": 35107181936,
+            "path": "qwen3_5_moe",
+        },
+        "kaggle_handle": "kaggle://keras/qwen3-5-moe/keras/qwen3_5_moe_35b_a3b/1",
+    },
+    # Qwen3.6 MoE Model Presets
+    "qwen3_6_moe_35b_a3b": {
+        "metadata": {
+            "description": (
+                "35 billion total parameter Qwen3.6 MoE instruction-tuned "
+                "model with ~3 billion active parameters per token. "
+                "Features a 3:1 hybrid attention stack (GatedDeltaNet "
+                "linear attention and full attention) with sparse "
+                "Mixture-of-Experts feedforward. Optimized for fast "
+                "inference and extended context lengths."
+            ),
+            "params": 35107181936,
+            "path": "qwen3_5_moe",
+        },
+        "kaggle_handle": "kaggle://keras/qwen3-6-moe/keras/qwen3_6_moe_35b_a3b/1",
+    },
+}