[MODEL] support hy_v3 and ministral3

README.md

@@ -21,6 +21,7 @@
 
 ## Latest News
 
+* 05/25/2026 7.1.0-dev `main`: ✨ Added `hy_3` and `ministral3` model support
 * 05/25/2026 7.1.0-dev `main`: ✨ Added `hunyuan_v1_dense` and `hunyuan_v1_moe` model support
 * 05/21/2026 7.1.0-dev `main`: ✨ Added `nemotron_labs_diffusion` model support
 * 05/20/2026 7.1.0-dev `main`: ✨ Added `interns1`, `ovis2_5`, `ovis2_6_moe` and `ovis2_6_next` model support
@@ -251,7 +252,7 @@ Selected public references where teams or companies explicitly mention GPT-QMode
 
 | Model                    |   |                                 |  |                  |  |                                 |  |                        |   |
 |--------------------------|---|---------------------------------|--|------------------|--|---------------------------------|--|------------------------|---|
-| Apertus                  | ✅ | EXAONE 3/4                      | ✅ | Dots1            | ✅ | Mistral3                        | ✅ | Qwen 2/3/3.5 (Next/MoE) | ✅ |
+| Apertus                  | ✅ | EXAONE 3/4                      | ✅ | Dots1            | ✅ | Mistral3 / Ministral3           | ✅ | Qwen 2/3/3.5 (Next/MoE) | ✅ |
 | Baichuan                 | ✅ | Falcon (H1 / Mamba)             | ✅ | InternLM 1/2/2.5 | ✅ | Mixtral                         | ✅ | Qwen 2/2.5/3 VL        | ✅ |
 | Bloom                    | ✅ | FastVLM                         | ✅ | Kimi K2          | ✅ | MobileLLM                       | ✅ | Qwen 2.5/3 Omni        | ✅ |
 | ChatGLM                  | ✅ | Gemma 1-4 / 3n                  | ✅ | Klear            | ✅ | MOSS                            | ✅ | RefinedWeb             | ✅ |
@@ -266,7 +267,7 @@ Selected public references where teams or companies explicitly mention GPT-QMode
 | XVERSE                   | ✅ | Brumby                          | ✅ | Hymba            | ✅ | Mistral                         | ✅ | Qwen 1/2/3/3.5         | ✅ |
 | MiniMax M2               | ✅ | AfMoE                           | ✅ | Bailing-MoE      | ✅ | LFM2-MoE                        | ✅ | Marin                  | ✅ |
 | InternVL Chat            | ✅ | Laguna                          | ✅ | Mimo / Mimo V2   | ✅ | Zamba / Zamba2                  | ✅ | Intern S1              | ✅ |
-| HunYuan V1 Dense / MoE   | ✅ |                           |  |    |  |                  |  |                |   |
+| HunYuan V1 Dense / MoE   | ✅ | HY-V3                           | ✅ |    |  |                  |  |                |   |
 
 Prism Bonsai GGUF checkpoints are supported for inference only through GPT-QModel's native GGUF path and internal GGUF runtime. Bonsai checkpoints load through the normal model path or repo argument and do not require the external `gguf` package. Prism model quantization is not included.
 

gptqmodel/models/auto.py

@@ -117,6 +117,7 @@
 from .definitions.hrm_text import HrmTextQModel  # noqa: E402
 from .definitions.hunyuan_v1_dense import HunYuanDenseV1QModel  # noqa: E402
 from .definitions.hunyuan_v1_moe import HunYuanMoEV1QModel  # noqa: E402
+from .definitions.hy_v3 import HYV3QModel  # noqa: E402
 from .definitions.hymba import HymbaQModel  # noqa: E402
 from .definitions.instella import InstellaQModel  # noqa: E402
 from .definitions.internlm import InternLMQModel  # noqa: E402
@@ -140,6 +141,7 @@
 from .definitions.minicpmv import MiniCPMVQModel  # noqa: E402
 from .definitions.minicpmv_4_6 import MiniCPMV4_6QModel  # noqa: E402
 from .definitions.minimax_m2 import MiniMaxM2GPTQ  # noqa: E402
+from .definitions.ministral3 import Ministral3GPTQ  # noqa: E402
 from .definitions.mistral3 import Mistral3GPTQ
 from .definitions.mixtral import MixtralQModel  # noqa: E402
 from .definitions.mllama import MLlamaQModel, MLlamaTextQModel  # noqa: E402
@@ -237,6 +239,7 @@
     "hrm_text": HrmTextQModel,
     "hunyuan_v1_dense": HunYuanDenseV1QModel,
     "hunyuan_v1_moe": HunYuanMoEV1QModel,
+    "hy_v3": HYV3QModel,
     "qwen": QwenQModel,
     "mistral": LlamaQModel, # 100% llama clone
     "yi": LlamaQModel, # 100% llama clone
@@ -272,6 +275,7 @@
     "minicpmv4_6": MiniCPMV4_6QModel,
     "minimax": MiniMaxM2GPTQ,
     "minimax_m2": MiniMaxM2GPTQ,
+    "ministral3": Ministral3GPTQ,
     "qwen2_moe": Qwen2MoeQModel,
     "qwen3_moe": Qwen3MoeQModel,
     "qwen3_next": Qwen3NextGPTQ,

gptqmodel/models/definitions/__init__.py

@@ -44,6 +44,7 @@
 from .hrm_text import HrmTextQModel
 from .hunyuan_v1_dense import HunYuanDenseV1QModel
 from .hunyuan_v1_moe import HunYuanMoEV1QModel
+from .hy_v3 import HYV3QModel
 from .hymba import HymbaQModel
 from .instella import InstellaQModel
 from .internlm import InternLMQModel
@@ -57,6 +58,7 @@
 from .minicpmv import MiniCPMVQModel
 from .minicpmv_4_6 import MiniCPMV4_6QModel
 from .minimax_m2 import MiniMaxM2GPTQ
+from .ministral3 import Ministral3GPTQ
 from .mimo_v2 import MimoV2QModel
 from .mixtral import MixtralQModel
 from .mllama import MLlamaQModel, MLlamaTextQModel

gptqmodel/models/definitions/hy_v3.py

@@ -0,0 +1,46 @@
+# SPDX-FileCopyrightText: 2026 ModelCloud.ai
+# SPDX-License-Identifier: Apache-2.0
+
+from ..base import BaseQModel
+from ..moe_lifecycle import GateUpDownMoELifecycleHooks
+
+
+class HYV3QModel(BaseQModel):
+    # HYV3 uses a dense first MLP layer and sparse MoE layers after it.
+    layer_modules_strict = False
+    dynamic_expert_index = "num_experts"
+
+    pre_lm_head_norm_module = "model.norm"
+
+    awq_scale_optimize_shape_dependent_modules = ["self_attn.o_proj"]
+
+    moe_lifecycle_hooks = GateUpDownMoELifecycleHooks()
+
+    module_tree = [
+        "model",
+        "layers",
+        "#",
+        {
+            "input_layernorm": ("input_layernorm:!",),
+            "self_attn": (
+                "q_norm:!",
+                "k_norm:!",
+                "q_proj:0",
+                "k_proj:0",
+                "v_proj:0",
+                "o_proj:1",
+            ),
+            "post_attention_layernorm": ("post_attention_layernorm:!",),
+            "mlp:moe": {
+                "gate": ("gate:!",),
+                "experts": {
+                    "#": ("gate_proj:0", "up_proj:0", "down_proj:1"),
+                },
+                "shared_experts": ("gate_proj:0", "up_proj:0", "down_proj:1"),
+                "": ("gate_proj:0", "up_proj:0", "down_proj:1"),
+            },
+        },
+    ]
+
+
+__all__ = ["HYV3QModel"]

gptqmodel/models/definitions/ministral3.py

@@ -0,0 +1,23 @@
+# SPDX-FileCopyrightText: 2026 ModelCloud.ai
+# SPDX-License-Identifier: Apache-2.0
+
+from ..base import BaseQModel
+
+
+class Ministral3GPTQ(BaseQModel):
+    pre_lm_head_norm_module = "model.norm"
+
+    module_tree = [
+        "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"),
+        },
+    ]
+
+
+__all__ = ["Ministral3GPTQ"]

tests/models/test_hunyuan_v1_dense.py

@@ -6,7 +6,7 @@
 from model_test import ModelTest
 
 
-class TestNemotronUltra(ModelTest):
+class TestHunyuanV1Dense(ModelTest):
     NATIVE_MODEL_ID = "/monster/data/model/HY-MT1.5-1.8B" # tencent/HY-MT1.5-1.8B
     EVAL_TASKS_SLOW = {
         "arc_challenge": {
@@ -24,6 +24,5 @@ class TestNemotronUltra(ModelTest):
     }
     EVAL_TASKS_FAST = ModelTest.derive_fast_eval_tasks(EVAL_TASKS_SLOW)
 
-    def test_nemotron_ultra(self):
-        # self.quantize_and_evaluate()
-        print(self.evaluate_model(self.SAVE_PATH))
+    def test_hunyuan_v1_dense(self):
+        self.quantize_and_evaluate()

tests/models/test_hunyuan_v1_moe.py

@@ -6,7 +6,7 @@
 from model_test import ModelTest
 
 
-class TestNemotronUltra(ModelTest):
+class TestHunyuanV1MoE(ModelTest):
     NATIVE_MODEL_ID = "/monster/data/model/Hunyuan-A13B-Instruct" # tencent/Hunyuan-A13B-Instruct
     EVAL_TASKS_SLOW = {
         "arc_challenge": {
@@ -24,5 +24,5 @@ class TestNemotronUltra(ModelTest):
     }
     EVAL_TASKS_FAST = ModelTest.derive_fast_eval_tasks(EVAL_TASKS_SLOW)
 
-    def test_nemotron_ultra(self):
+    def test_hunyuan_v1_moe(self):
         self.quantize_and_evaluate()

tests/models/test_hy_v3.py

@@ -0,0 +1,21 @@
+# SPDX-FileCopyrightText: 2024-2025 ModelCloud.ai
+# SPDX-FileCopyrightText: 2024-2025 qubitium@modelcloud.ai
+# SPDX-License-Identifier: Apache-2.0
+# Contact: qubitium@modelcloud.ai, x.com/qubitium
+
+from model_test import ModelTest
+
+
+class TestHyV3(ModelTest):
+    NATIVE_MODEL_ID = "/monster/data/model/Hy-MT2-30B-A3B" # tencent/Hy-MT2-30B-A3B
+    EVAL_TASKS_SLOW = {
+        "arc_challenge": {
+            "chat_template": True,
+            "acc": {"value": 0.5324, "floor_pct": 0.04},
+            "acc_norm": {"value": 0.5341, "floor_pct": 0.04},
+        },
+    }
+    EVAL_TASKS_FAST = ModelTest.derive_fast_eval_tasks(EVAL_TASKS_SLOW)
+
+    def test_hy_v3(self):
+        self.quantize_and_evaluate()

tests/models/test_ministral3.py

@@ -0,0 +1,29 @@
+# SPDX-FileCopyrightText: 2024-2025 ModelCloud.ai
+# SPDX-FileCopyrightText: 2024-2025 qubitium@modelcloud.ai
+# SPDX-License-Identifier: Apache-2.0
+# Contact: qubitium@modelcloud.ai, x.com/qubitium
+
+from model_test import ModelTest
+
+
+class TestMinistral3(ModelTest):
+    NATIVE_MODEL_ID = "/monster/data/model/Ministral-3-3B-Instruct-2512-TextOnly-BF16" # "Aratako/Ministral-3-3B-Instruct-2512-TextOnly"
+    NATIVE_ARC_CHALLENGE_ACC = 0.5870
+    NATIVE_ARC_CHALLENGE_ACC_NORM = 0.6032
+    NATIVE_ARC_CHALLENGE_ACC_SLOW = NATIVE_ARC_CHALLENGE_ACC
+    NATIVE_ARC_CHALLENGE_ACC_NORM_SLOW = NATIVE_ARC_CHALLENGE_ACC_NORM
+    NATIVE_ARC_CHALLENGE_ACC_FAST = NATIVE_ARC_CHALLENGE_ACC_SLOW
+    NATIVE_ARC_CHALLENGE_ACC_NORM_FAST = NATIVE_ARC_CHALLENGE_ACC_NORM_SLOW
+    TRUST_REMOTE_CODE = False
+    EVAL_BATCH_SIZE = 6
+    EVAL_TASKS_SLOW = {
+        "arc_challenge": {
+            "chat_template": False,
+            "acc": {"value": NATIVE_ARC_CHALLENGE_ACC},
+            "acc_norm": {"value": NATIVE_ARC_CHALLENGE_ACC_NORM},
+        },
+    }
+    EVAL_TASKS_FAST = ModelTest.derive_fast_eval_tasks(EVAL_TASKS_SLOW)
+
+    def test_ministral3(self):
+        self.quantize_and_evaluate()

tests/test_hy_v3_support.py

@@ -0,0 +1,43 @@
+from types import SimpleNamespace
+
+from gptqmodel.models import auto
+from gptqmodel.models.definitions.hy_v3 import HYV3QModel
+
+def test_hy_v3_model_type_selects_definition(monkeypatch):
+    fake_config = SimpleNamespace(model_type="hy_v3")
+
+    monkeypatch.setattr(auto, "resolve_trust_remote_code", lambda path, trust_remote_code=False: trust_remote_code)
+    monkeypatch.setattr(auto.AutoConfig, "from_pretrained", lambda *args, **kwargs: fake_config)
+
+    assert auto.check_and_get_model_definition("/tmp/hy_v3") is HYV3QModel
+
+
+def test_hy_v3_module_tree_expands_dense_and_sparse_moe_paths():
+    layer_modules = HYV3QModel.simple_layer_modules(
+        model_config=SimpleNamespace(num_experts=3),
+        quantize_config=SimpleNamespace(dynamic=None),
+    )
+    flat_modules = {name for block in layer_modules for name in block}
+    first_expert_block = next(i for i, block in enumerate(layer_modules) if "mlp.experts.0.gate_proj" in block)
+    shared_block = next(i for i, block in enumerate(layer_modules) if "mlp.shared_experts.gate_proj" in block)
+
+    assert HYV3QModel.layer_modules_strict is False
+    assert HYV3QModel.dynamic_expert_index == "num_experts"
+    assert "self_attn.q_proj" in flat_modules
+    assert "self_attn.k_proj" in flat_modules
+    assert "self_attn.v_proj" in flat_modules
+    assert "self_attn.o_proj" in flat_modules
+    assert "self_attn.q_norm" not in flat_modules
+    assert "self_attn.k_norm" not in flat_modules
+    assert "mlp.gate_proj" in flat_modules
+    assert "mlp.up_proj" in flat_modules
+    assert "mlp.down_proj" in flat_modules
+    assert "mlp.shared_experts.gate_proj" in flat_modules
+    assert "mlp.shared_experts.up_proj" in flat_modules
+    assert "mlp.shared_experts.down_proj" in flat_modules
+    assert "mlp.experts.0.gate_proj" in flat_modules
+    assert "mlp.experts.1.up_proj" in flat_modules
+    assert "mlp.experts.2.down_proj" in flat_modules
+    assert "mlp.gate" not in flat_modules
+    assert "mlp.e_score_correction_bias" not in flat_modules
+    assert first_expert_block < shared_block

tests/test_ministral3_support.py

@@ -0,0 +1,31 @@
+from types import SimpleNamespace
+
+from gptqmodel.models import auto
+from gptqmodel.models.definitions.ministral3 import Ministral3GPTQ
+
+
+def test_ministral3_model_type_selects_definition(monkeypatch):
+    fake_config = SimpleNamespace(model_type="ministral3")
+
+    monkeypatch.setattr(auto, "resolve_trust_remote_code", lambda path, trust_remote_code=False: trust_remote_code)
+    monkeypatch.setattr(auto.AutoConfig, "from_pretrained", lambda *args, **kwargs: fake_config)
+
+    assert auto.check_and_get_model_definition("/tmp/ministral3") is Ministral3GPTQ
+
+
+def test_ministral3_module_tree_matches_text_only_layout():
+    layer_modules = Ministral3GPTQ.simple_layer_modules(
+        model_config=SimpleNamespace(),
+        quantize_config=SimpleNamespace(dynamic=None),
+    )
+    flat_modules = {name for block in layer_modules for name in block}
+
+    assert Ministral3GPTQ.module_tree[:3] == ["model", "layers", "#"]
+    assert "self_attn.q_proj" in flat_modules
+    assert "self_attn.k_proj" in flat_modules
+    assert "self_attn.v_proj" in flat_modules
+    assert "self_attn.o_proj" in flat_modules
+    assert "mlp.gate_proj" in flat_modules
+    assert "mlp.up_proj" in flat_modules
+    assert "mlp.down_proj" in flat_modules
+