ModelCloud · ZX-ModelCloud · Jul 7, 2026
diff --git a/README.md b/README.md
@@ -21,6 +21,7 @@
 
 ## Latest News
 
+* 07/07/2026 7.1.0-dev `main`: ✨ Added `deepseek_vl` model support
 * 07/06/2026 7.1.0-dev `main`: ✨ Added `deepseek_ocr2` model support
 * 04/07/2026 7.1.0-dev `main`: ✨ Added `deepseek_vl_v2` model support
 * 02/07/2026 7.1.0-dev `main`: ✨ Added `lfm2` and `lfm2_vl` model support
@@ -266,7 +267,7 @@ Selected public references where teams or companies explicitly mention GPT-QMode
 | DBRX Converted                | ✅ | GPT-2                           | ✅ | Llama 3.2 VL     | ✅ | Nemotron Ultra / Labs-Diffusion               | ✅ | TeleChat2              | ✅ |
 | Deci                          | ✅ | GPT-J                           | ✅ | Llama 4          | ✅ | OPT                             | ✅ | Trinity                | ✅ |
 | DeepSeek-V2/V3/V4/R1          | ✅ | GPT-OSS                         | ✅ | LongCat Flash    | ✅ | OLMo2 / LLaDA2                  | ✅ | Yi                     | ✅ |
-| DeepSeek-V2 Lite / VL2 / OCR2 | ✅ | Granite / Granite MoE           | ✅ | LongLLaMA        | ✅ | Ovis 1.6/2/2.5/2.6 MoE/2.6 Next | ✅ | Seed-OSS               | ✅ |
+| DeepSeek-V2 Lite / VL / VL2 / OCR2 | ✅ | Granite / Granite MoE           | ✅ | LongLLaMA        | ✅ | Ovis 1.6/2/2.5/2.6 MoE/2.6 Next | ✅ | Seed-OSS               | ✅ |
 | Dream                         | ✅ | GRIN-MoE                        | ✅ | Instella         | ✅ | Phi 1-4                         | ✅ | Voxtral                | ✅ |
 | ERNIE 4.5 / MoE / VL MoE      | ✅ | GLM 4/4V/4.5V/4.6V/5/5.1/OCR/ASR | ✅ | GLM4 MoE / Lite / 4.5V MoE | ✅ | MiniCPM 3/O/V/V 4_6             | ✅ | PanGu-α                | ✅ |
 | XVERSE                        | ✅ | Brumby                          | ✅ | Hymba            | ✅ | Mistral                         | ✅ | Qwen 1/2/3/3.5         | ✅ |

diff --git a/gptqmodel/models/auto.py b/gptqmodel/models/auto.py
@@ -87,6 +87,7 @@
 from .definitions.deepseek_v3 import DeepSeekV3QModel  # noqa: E402
 from .definitions.deepseek_v4 import DeepSeekV4QModel  # noqa: E402
 from .definitions.deepseek_ocr2 import DeepSeekOCR2QModel  # noqa: E402
+from .definitions.deepseek_vl import DeepSeekVLQModel  # noqa: E402
 from .definitions.deepseek_vl_v2 import DeepSeekVLV2QModel  # noqa: E402
 from .definitions.dots1 import Dots1QModel  # noqa: E402
 from .definitions.dream import DreamQModel  # noqa: E402
@@ -301,6 +302,7 @@
     "deepseek_v3": DeepSeekV3QModel,
     "deepseek_v4": DeepSeekV4QModel,
     "deepseek_ocr2": DeepSeekOCR2QModel,
+    "deepseek_vl": DeepSeekVLQModel,
     "deepseek_vl_v2": DeepSeekVLV2QModel,
     "dots1": Dots1QModel,
     "exaone": ExaOneQModel,

diff --git a/gptqmodel/models/definitions/__init__.py b/gptqmodel/models/definitions/__init__.py
@@ -21,6 +21,7 @@
 from .deepseek_v3 import DeepSeekV3QModel
 from .deepseek_v4 import DeepSeekV4QModel
 from .deepseek_ocr2 import DeepSeekOCR2QModel
+from .deepseek_vl import DeepSeekVLQModel
 from .deepseek_vl_v2 import DeepSeekVLV2QModel
 from .dots1 import Dots1QModel
 from .dream import DreamQModel

diff --git a/gptqmodel/models/definitions/deepseek_vl.py b/gptqmodel/models/definitions/deepseek_vl.py
@@ -0,0 +1,144 @@
+# SPDX-FileCopyrightText: 2026 ModelCloud.ai
+# SPDX-FileCopyrightText: 2026 qubitium@modelcloud.ai
+# SPDX-License-Identifier: Apache-2.0
+# Contact: qubitium@modelcloud.ai, x.com/qubitium
+
+from types import SimpleNamespace
+from typing import Any, Dict
+
+import torch
+from transformers import AutoModelForImageTextToText, AutoProcessor, ProcessorMixin
+
+from ...utils.calibration import batched
+from ...utils.model import MODALITY, move_to
+from ...utils.offload import offload_to_disk
+from .._const import CPU
+from ..base import BaseQModel
+
+
+class DeepSeekVLQModel(BaseQModel):
+    loader = AutoModelForImageTextToText
+    require_load_processor = True
+    support_batch_quantize = False
+
+    modality = [MODALITY.TEXT, MODALITY.IMAGE_TO_TEXT]
+
+    pre_lm_head_norm_module = "model.language_model.norm"
+    rotary_embedding = "model.language_model.rotary_emb"
+    # layer_modules_strict = False
+
+    # HF_CONVERSION_MAP_REVERSED = (
+    #     # DeepSeek-VL mounts `SiglipVisionModel` at `model.vision_model`.
+    #     # Runtime shell paths expose `model.vision_model.*`, while the original
+    #     # checkpoint keeps SigLIP's base prefix under `model.vision_model.vision_model.*`.
+    #     SimpleNamespace(
+    #         source_patterns=[r"^model\.vision_model\.(?!vision_model\.)(.+)$"],
+    #         target_patterns=[r"^model.vision_model.vision_model.\1"],
+    #         operations=[],
+    #     ),
+    # )
+
+    module_tree = [
+        "model",
+        "language_model",
+        "layers",
+        "#",
+        {
+            "input_layernorm": ("input_layernorm:!",),
+            "self_attn": ("q_proj:0", "k_proj:0", "v_proj:0", "o_proj:1"),
+            "post_attention_layernorm": ("post_attention_layernorm:!",),
+            "mlp": ("gate_proj:0", "up_proj:0", "down_proj:1"),
+        },
+    ]
+
+    def pre_quantize_generate_hook_start(self):
+        core_model = self.model.model
+        language_model = core_model.language_model
+        self.shell_module_materialize(language_model.embed_tokens, self.quantize_config.device)
+        self.shell_module_materialize(language_model.norm, self.quantize_config.device)
+        self.shell_module_materialize(language_model.rotary_emb, self.quantize_config.device)
+        self.shell_module_materialize(core_model.vision_model, self.quantize_config.device)
+        self.shell_module_materialize(core_model.aligner, self.quantize_config.device)
+
+    def pre_quantize_generate_hook_end(self):
+        core_model = self.model.model
+        language_model = core_model.language_model
+        if self.quantize_config.offload_to_disk:
+            offload_to_disk(
+                model=language_model,
+                module=language_model.embed_tokens,
+                disk_path=self.quantize_config.offload_to_disk_path,
+            )
+            offload_to_disk(
+                model=language_model,
+                module=language_model.norm,
+                disk_path=self.quantize_config.offload_to_disk_path,
+            )
+            offload_to_disk(
+                model=language_model,
+                module=language_model.rotary_emb,
+                disk_path=self.quantize_config.offload_to_disk_path,
+            )
+            offload_to_disk(
+                model=core_model,
+                module=core_model.vision_model,
+                disk_path=self.quantize_config.offload_to_disk_path,
+            )
+            offload_to_disk(
+                model=core_model,
+                module=core_model.aligner,
+                disk_path=self.quantize_config.offload_to_disk_path,
+            )
+            return
+
+        language_model.embed_tokens = move_to(language_model.embed_tokens, device=CPU)
+        language_model.norm = move_to(language_model.norm, device=CPU)
+        language_model.rotary_emb = move_to(language_model.rotary_emb, device=CPU)
+        core_model.vision_model = move_to(core_model.vision_model, device=CPU)
+        core_model.aligner = move_to(core_model.aligner, device=CPU)
+
+    def preprocess_dataset(self, sample: Dict) -> Dict:
+        return sample
+
+    def load_processor(self) -> ProcessorMixin:
+        return AutoProcessor.from_pretrained(self.model_local_path, trust_remote_code=False)
+
+    @classmethod
+    def prepare_inputs_for_conversations(
+        cls,
+        processor: ProcessorMixin,
+        conversations: list[dict] | list[list[dict]],
+    ):
+        return processor.apply_chat_template(
+            conversations,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+        )
+
+    def prepare_dataset(self, calibration_dataset, batch_size: int = 1, **kwargs):
+        del kwargs
+        processor = self.load_processor()
+        calib_data = []
+        for batch in batched(calibration_dataset, batch_size, process_func=self.preprocess_dataset):
+            calib_data.append(self.prepare_inputs_for_conversations(processor, batch))
+        del processor
+        return calib_data
+
+    def move_input_capture_example(self, example: Dict[str, Any], data_device: torch.device) -> Dict[str, Any]:
+        example = super().move_input_capture_example(example, data_device)
+        pixel_values = example.get("pixel_values")
+        if torch.is_tensor(pixel_values):
+            vision_model = self.model.model.vision_model
+            first_parameter = next(vision_model.parameters(), None)
+            vision_device = getattr(first_parameter, "device", pixel_values.device)
+            vision_dtype = getattr(vision_model, "dtype", None)
+            if not isinstance(vision_dtype, torch.dtype):
+                vision_dtype = getattr(first_parameter, "dtype", None)
+            if isinstance(vision_dtype, torch.dtype):
+                example["pixel_values"] = pixel_values.to(device=vision_device, dtype=vision_dtype)
+        return example
+
+
+__all__ = ["DeepSeekVLQModel"]
diff --git a/tests/models/ovis/image_to_test_dataset.py b/tests/models/ovis/image_to_test_dataset.py
@@ -5,6 +5,7 @@
 from gptqmodel.models.definitions.base_qwen2_5_omni import BaseQwen2_5_OmniGPTQ
 from gptqmodel.models.definitions.base_qwen2_vl import BaseQwen2VLGPTQ
 from gptqmodel.models.definitions.deepseek_ocr2 import DeepSeekOCR2QModel
+from gptqmodel.models.definitions.deepseek_vl import DeepSeekVLQModel
 from gptqmodel.models.definitions.deepseek_vl_v2 import DeepSeekVLV2QModel
 from gptqmodel.models.definitions.ernie4_5_vl_moe import Ernie4_5_VLMoeQModel
 from gptqmodel.models.definitions.interns1 import InternS1QModel
@@ -72,6 +73,24 @@ def format_deepseek_vl_v2_dataset(image, assistant):
     ]
 
 
+def format_deepseek_vl_dataset(image, assistant):
+    return [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image", "url": image},
+                {"type": "text", "text": "generate a caption for this image"},
+            ],
+        },
+        {
+            "role": "assistant",
+            "content": [
+                {"type": "text", "text": assistant},
+            ],
+        },
+    ]
+
+
 def format_deepseek_ocr2_dataset(image, assistant):
     del assistant
     return {
@@ -116,6 +135,10 @@ def prepare_deepseek_vl_v2_dataset(n_sample: int = 20) -> list[list[dict]]:
     return prepare_dataset(format_deepseek_vl_v2_dataset, n_sample=n_sample)
 
 
+def prepare_deepseek_vl_dataset(n_sample: int = 20) -> list[list[dict]]:
+    return prepare_dataset(format_deepseek_vl_dataset, n_sample=n_sample)
+
+
 def prepare_deepseek_ocr2_dataset(n_sample: int = 20) -> list[dict]:
     return prepare_dataset(format_deepseek_ocr2_dataset, n_sample=n_sample)
 
@@ -152,6 +175,9 @@ def get_calib_dataset(model):
     if isinstance(model, DeepSeekVLV2QModel):
         return prepare_deepseek_vl_v2_dataset(n_sample=20)
 
+    if isinstance(model, DeepSeekVLQModel):
+        return prepare_deepseek_vl_dataset(n_sample=20)
+
     if isinstance(model, DeepSeekOCR2QModel):
         return prepare_deepseek_ocr2_dataset(n_sample=20)