diff --git a/mlx_lm/models/minimax_m3_vl.py b/mlx_lm/models/minimax_m3_vl.py
new file mode 100644
index 000000000..52bfccae6
--- /dev/null
+++ b/mlx_lm/models/minimax_m3_vl.py
@@ -0,0 +1,429 @@
+# Copyright © 2025 Apple Inc.
+
+from dataclasses import dataclass
+from functools import partial
+from typing import Any, List, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx.nn.layers.distributed import shard_inplace, shard_linear, sum_gradients
+
+from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention
+from .switch_layers import SwitchGLU
+
+
+@partial(mx.compile, shapeless=True)
+def _swiglu_oai(gate, up, alpha, limit):
+    gate = mx.clip(gate, -limit, limit)
+    up = mx.clip(up, -limit, limit)
+    return (up + 1.0) * (gate * mx.sigmoid(gate * alpha))
+
+
+class SwigluOAI(nn.Module):
+    def __init__(self, alpha: float = 1.702, limit: float = 7.0):
+        super().__init__()
+        self._alpha = alpha
+        self._limit = limit
+
+    def __call__(self, x, gate):
+        return _swiglu_oai(gate, x, self._alpha, self._limit)
+
+
+class GemmaRMSNorm(nn.Module):
+    def __init__(self, dims: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = mx.zeros((dims,))
+        self.eps = eps
+
+    def _extra_repr(self):
+        return f"{self.weight.shape[0]}, eps={self.eps}"
+
+    def __call__(self, x):
+        t = x.dtype
+        x = x.astype(mx.float32)
+        n = mx.rsqrt(x.square().mean(-1, keepdims=True) + self.eps)
+        return ((x * n) * (1.0 + self.weight)).astype(t)
+
+
+@dataclass
+class TextArgs(BaseModelArgs):
+    model_type: str = ""
+    hidden_size: int = 6144
+    intermediate_size: int = 3072
+    dense_intermediate_size: int = 12288
+    shared_intermediate_size: int = 3072
+    num_attention_heads: int = 64
+    num_key_value_heads: int = 4
+    num_hidden_layers: int = 60
+    num_local_experts: int = 128
+    num_experts_per_tok: int = 4
+    max_position_embeddings: int = 1048576
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 5000000.0
+    head_dim: int = 128
+    rotary_dim: int = 64
+    partial_rotary_factor: float = 0.5
+    vocab_size: int = 200064
+    tie_word_embeddings: bool = False
+    scoring_func: str = "sigmoid"
+    routed_scaling_factor: float = 2.0
+    use_routing_bias: bool = True
+    use_qk_norm: bool = True
+    use_gemma_norm: bool = True
+    swiglu_alpha: float = 1.702
+    swiglu_limit: float = 7.0
+    moe_layer_freq: Optional[List[int]] = None
+    n_shared_experts: int = 1
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    text_config: Union[TextArgs, dict] = None
+    model_type: str = "minimax_m3_vl"
+
+    def __post_init__(self):
+        if isinstance(self.text_config, dict):
+            self.text_config = TextArgs.from_dict(self.text_config)
+        if self.text_config is None:
+            self.text_config = TextArgs()
+
+
+class DenseMLP(nn.Module):
+    def __init__(self, hidden: int, inter: int, alpha: float, limit: float):
+        super().__init__()
+        self.gate_proj = nn.Linear(hidden, inter, bias=False)
+        self.up_proj = nn.Linear(hidden, inter, bias=False)
+        self.down_proj = nn.Linear(inter, hidden, bias=False)
+        self._alpha = alpha
+        self._limit = limit
+
+    def __call__(self, x):
+        return self.down_proj(
+            _swiglu_oai(self.gate_proj(x), self.up_proj(x), self._alpha, self._limit)
+        )
+
+
+class SparseMoeBlock(nn.Module):
+    def __init__(self, args: TextArgs):
+        super().__init__()
+        self.num_experts_per_tok = args.num_experts_per_tok
+        self.routed_scaling_factor = args.routed_scaling_factor
+
+        self.gate = nn.Linear(args.hidden_size, args.num_local_experts, bias=False)
+        self.e_score_correction_bias = mx.zeros((args.num_local_experts,))
+
+        activation = SwigluOAI(args.swiglu_alpha, args.swiglu_limit)
+        self.switch_mlp = SwitchGLU(
+            args.hidden_size,
+            args.intermediate_size,
+            args.num_local_experts,
+            activation=activation,
+        )
+
+        self.shared_experts = DenseMLP(
+            args.hidden_size,
+            args.shared_intermediate_size,
+            args.swiglu_alpha,
+            args.swiglu_limit,
+        )
+
+        self.sharding_group = None
+
+    def __call__(self, x: mx.array) -> mx.array:
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+
+        ne = x.shape[:-1]
+        x_flat = x.reshape(-1, x.shape[-1])
+
+        shared_out = self.shared_experts(x_flat)
+
+        gates = self.gate(x_flat.astype(mx.float32))
+        scores = mx.sigmoid(gates)
+        orig_scores = scores
+        scores = scores + self.e_score_correction_bias
+
+        k = self.num_experts_per_tok
+        inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k]
+        scores = mx.take_along_axis(orig_scores, inds, axis=-1)
+        scores = scores / (mx.sum(scores, axis=-1, keepdims=True) + 1e-20)
+        scores = scores.astype(x_flat.dtype)
+
+        y = self.switch_mlp(x_flat, inds)
+        y = (y * scores[..., None]).sum(axis=-2)
+        y = y * self.routed_scaling_factor + shared_out
+
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group)
+
+        return y.reshape(*ne, -1)
+
+
+class Attention(nn.Module):
+    def __init__(self, args: TextArgs):
+        super().__init__()
+        self.hidden_dim = args.hidden_size
+        self.num_attention_heads = args.num_attention_heads
+        self.num_key_value_heads = args.num_key_value_heads
+        self.head_dim = args.head_dim
+        self.scale = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(
+            args.hidden_size, self.num_attention_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            args.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            args.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            self.num_attention_heads * self.head_dim, args.hidden_size, bias=False
+        )
+
+        self.use_qk_norm = args.use_qk_norm
+        if self.use_qk_norm:
+            NormClass = GemmaRMSNorm if args.use_gemma_norm else nn.RMSNorm
+            self.q_norm = NormClass(self.head_dim, eps=args.rms_norm_eps)
+            self.k_norm = NormClass(self.head_dim, eps=args.rms_norm_eps)
+
+        self.rope = nn.RoPE(args.rotary_dim, traditional=False, base=args.rope_theta)
+
+    def __call__(self, x, mask=None, cache=None):
+        B, L, _ = x.shape
+
+        q, k, v = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+        q = q.reshape(B, L, self.num_attention_heads, self.head_dim).transpose(
+            0, 2, 1, 3
+        )
+        k = k.reshape(B, L, self.num_key_value_heads, self.head_dim).transpose(
+            0, 2, 1, 3
+        )
+        v = v.reshape(B, L, self.num_key_value_heads, self.head_dim).transpose(
+            0, 2, 1, 3
+        )
+
+        if self.use_qk_norm:
+            q = self.q_norm(q)
+            k = self.k_norm(k)
+
+        if cache is not None:
+            q = self.rope(q, offset=cache.offset)
+            k = self.rope(k, offset=cache.offset)
+            k, v = cache.update_and_fetch(k, v)
+        else:
+            q = self.rope(q)
+            k = self.rope(k)
+
+        out = scaled_dot_product_attention(
+            q, k, v, cache=cache, scale=self.scale, mask=mask
+        )
+        out = out.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(out)
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, args: TextArgs, layer_idx: int):
+        super().__init__()
+        self.self_attn = Attention(args)
+
+        is_moe = bool(args.moe_layer_freq and args.moe_layer_freq[layer_idx])
+        if is_moe:
+            self.block_sparse_moe = SparseMoeBlock(args)
+        else:
+            self.mlp = DenseMLP(
+                args.hidden_size,
+                args.dense_intermediate_size,
+                args.swiglu_alpha,
+                args.swiglu_limit,
+            )
+        self.is_moe = is_moe
+
+        NormClass = GemmaRMSNorm if args.use_gemma_norm else nn.RMSNorm
+        self.input_layernorm = NormClass(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = NormClass(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+
+    def __call__(self, x, mask=None, cache=None):
+        r = x + self.self_attn(self.input_layernorm(x), mask, cache)
+        if self.is_moe:
+            r = r + self.block_sparse_moe(self.post_attention_layernorm(r))
+        else:
+            r = r + self.mlp(self.post_attention_layernorm(r))
+        return r
+
+
+class TextModel(nn.Module):
+    def __init__(self, args: TextArgs):
+        super().__init__()
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [DecoderLayer(args, i) for i in range(args.num_hidden_layers)]
+        NormClass = GemmaRMSNorm if args.use_gemma_norm else nn.RMSNorm
+        self.norm = NormClass(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(self, inputs, mask=None, cache=None):
+        h = self.embed_tokens(inputs)
+        if cache is None:
+            cache = [None] * len(self.layers)
+        mask = create_attention_mask(h, cache[0])
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+        return self.norm(h)
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextArgs):
+        super().__init__()
+        self.args = config
+        self.model = TextModel(config)
+        if not config.tie_word_embeddings:
+            self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+    def __call__(self, inputs, cache=None):
+        out = self.model(inputs, cache=cache)
+        if self.args.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return out
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.args = config
+        self.model_type = config.model_type
+        self.language_model = LanguageModel(config.text_config)
+
+    def __call__(self, inputs, cache=None):
+        return self.language_model(inputs, cache)
+
+    def sanitize(self, weights):
+        new = {}
+        for k, v in weights.items():
+            # Skip vision, MTP, and sparse attention indexer keys
+            if any(
+                k.startswith(p)
+                for p in (
+                    "visual.",
+                    "vision_tower.",
+                    "vision_model.",
+                    "multi_modal_projector.",
+                    "mm_projector.",
+                    "patch_merge_mlp.",
+                )
+            ):
+                continue
+            if any(
+                f in k
+                for f in (
+                    ".rotary_emb.",
+                    "index_q_proj",
+                    "index_k_proj",
+                    "index_q_norm",
+                    "index_k_norm",
+                    ".mtp_",
+                )
+            ):
+                continue
+
+            # Remap HF prefix → MLX prefix
+            nk = k
+            if nk.startswith("model."):
+                nk = "language_model." + nk
+            elif nk.startswith("lm_head."):
+                nk = "language_model." + nk
+
+            new[nk] = v
+        weights = new
+
+        # FP8 dequantize
+        dequanted = {}
+        for k, v in weights.items():
+            if "weight_scale_inv" in k:
+                wk = k.replace("_scale_inv", "")
+                dequanted[wk] = self._dequant_fp8(weights[wk], v)
+            elif k not in dequanted:
+                dequanted[k] = v
+        weights = dequanted
+
+        # Expert w1/w2/w3 stacking → SwitchGLU
+        args = self.args.text_config
+        for l in range(args.num_hidden_layers):
+            prefix = f"language_model.model.layers.{l}"
+            if f"{prefix}.block_sparse_moe.experts.0.w1.weight" not in weights:
+                continue
+            mapping = {"w1": "gate_proj", "w2": "down_proj", "w3": "up_proj"}
+            for orig, new_name in mapping.items():
+                if f"{prefix}.block_sparse_moe.experts.0.{orig}.weight" in weights:
+                    stacked = [
+                        weights.pop(
+                            f"{prefix}.block_sparse_moe.experts.{e}.{orig}.weight"
+                        )
+                        for e in range(args.num_local_experts)
+                    ]
+                    weights[
+                        f"{prefix}.block_sparse_moe.switch_mlp.{new_name}.weight"
+                    ] = mx.stack(stacked)
+
+        return weights
+
+    @staticmethod
+    def _dequant_fp8(weight, scale_inv):
+        weight = mx.from_fp8(weight, dtype=mx.bfloat16)
+        bs = 128
+        m, n = weight.shape
+        pad_b = (-m) % bs
+        pad_s = (-n) % bs
+        weight = mx.pad(weight, ((0, pad_b), (0, pad_s)))
+        weight = weight.reshape((m + pad_b) // bs, bs, (n + pad_s) // bs, bs)
+        weight = (weight * scale_inv[:, None, :, None]).reshape(
+            m + pad_b, n + pad_s
+        )
+        return weight[:m, :n].astype(mx.bfloat16)
+
+    def shard(self, group=None):
+        group = group or mx.distributed.init()
+        N = group.size()
+        for layer in self.language_model.model.layers:
+            attn = layer.self_attn
+            attn.q_proj = shard_linear(attn.q_proj, "all-to-sharded", group=group)
+            attn.k_proj = shard_linear(attn.k_proj, "all-to-sharded", group=group)
+            attn.v_proj = shard_linear(attn.v_proj, "all-to-sharded", group=group)
+            attn.o_proj = shard_linear(attn.o_proj, "sharded-to-all", group=group)
+            attn.num_attention_heads //= N
+            attn.num_key_value_heads //= N
+
+            if layer.is_moe:
+                moe = layer.block_sparse_moe
+                shard_inplace(
+                    moe.switch_mlp.gate_proj, "all-to-sharded", group=group
+                )
+                shard_inplace(
+                    moe.switch_mlp.down_proj, "sharded-to-all", group=group
+                )
+                shard_inplace(
+                    moe.switch_mlp.up_proj, "all-to-sharded", group=group
+                )
+                moe.sharding_group = group
+
+    @property
+    def layers(self):
+        return self.language_model.model.layers
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "e_score_correction_bias" not in k
+
+        return predicate
+
+    @property
+    def quant_predicate(self):
+        def predicate(path, _):
+            if path.endswith("block_sparse_moe.gate"):
+                return {"group_size": 64, "bits": 8}
+            return True
+
+        return predicate