diff --git a/docs/models/vlm/README.md b/docs/models/vlm/README.md
index 86ab3cb955..71f9d1dd48 100644
--- a/docs/models/vlm/README.md
+++ b/docs/models/vlm/README.md
@@ -9,6 +9,7 @@ Megatron Bridge supports the following VLM families:
 | Model | Documentation | Description |
 |-------|---------------|-------------|
 | **Gemma 3 VL** | [gemma3-vl.md](gemma3-vl.md) | Google Gemma 3 Vision Language model |
+| **Ministral 3** | [ministral3.md](ministral3.md) | Ministral 3 Vision Language model |
 | **Nemotron Nano V2 VL** | [nemotron-nano-v2-vl.md](nemotron-nano-v2-vl.md) | NVIDIA Nemotron Nano V2 Vision Language model |
 | **Qwen2.5 VL** | [qwen2.5-vl.md](qwen2.5-vl.md) | Alibaba Cloud Qwen2.5 Vision Language model |
 | **Qwen3 VL** | [qwen3-vl.md](qwen3-vl.md) | Alibaba Cloud Qwen3 Vision Language model |
diff --git a/examples/conversion/hf_to_megatron_generate_vlm.py b/examples/conversion/hf_to_megatron_generate_vlm.py
index 71583afb88..5658b7dfef 100644
--- a/examples/conversion/hf_to_megatron_generate_vlm.py
+++ b/examples/conversion/hf_to_megatron_generate_vlm.py
@@ -115,7 +115,13 @@ def vlm_forward_step(data_iterator, model, **kwargs) -> torch.Tensor:
     def loss_func(x, **kwargs):
         return x
 
-    return model(**forward_args), loss_func
+    model_output = model(**forward_args)
+    if isinstance(model_output, tuple):
+        output_tensor, _ = model_output
+    else:
+        output_tensor = model_output
+
+    return output_tensor, loss_func
 
 
 def load_image(image_path: str) -> Image.Image:
diff --git a/examples/models/vlm/ministral3/README.md b/examples/models/vlm/ministral3/README.md
new file mode 100644
index 0000000000..16b3b7dbe7
--- /dev/null
+++ b/examples/models/vlm/ministral3/README.md
@@ -0,0 +1,190 @@
+# Ministral 3 - Vision Language Model
+
+[Mistral AI's Ministral 3](https://huggingface.co/collections/mistralai/ministral-3) is a family of edge-optimized vision-language models designed for deployment across various hardware configurations. The Ministral 3 architecture combines a powerful language model with a vision encoder for multimodal understanding.
+
+Ministral 3 models support multimodal tasks including image captioning, visual question answering, OCR, and general vision-language understanding. Despite their compact size, these models deliver strong performance for on-device and edge deployment scenarios.
+
+Ministral family models are supported via the Bridge system with auto-detected configuration and weight mapping.
+
+```{important}
+Please upgrade to `transformers` v5 and upgrade `mistral-common` in order to use the Ministral 3 models.
+```
+
+## Available Models
+
+### Vision-Language Models
+- **Ministral 3 3B** (`mistralai/Ministral-3-3B-Base-2512`): 3.4B parameter vision-language model
+  - 26 layers, 3072 hidden size
+  - 32 attention heads, 8 query groups (GQA)
+  - Vision encoder: ~0.4B parameters
+  - Recommended: 1 node, 8 GPUs
+
+- **Ministral 3 8B** (`mistralai/Ministral-3-8B-Base-2512`): 8.4B parameter vision-language model
+  - 34 layers, 4096 hidden size
+  - 32 attention heads, 8 query groups (GQA)
+  - Vision encoder: ~0.4B parameters
+  - Recommended: 1 node, 8 GPUs
+
+- **Ministral 3 14B** (`mistralai/Ministral-3-14B-Base-2512`): ~14B parameter vision-language model
+  - 40 layers, 5120 hidden size
+  - 32 attention heads, 8 query groups (GQA)
+  - Vision encoder: ~0.4B parameters
+  - Recommended: 1 node, 8 GPUs
+
+All models support extended context lengths up to 256K tokens using YaRN RoPE scaling.
+
+## Model Architecture Features
+
+Ministral 3 combines efficient language modeling with multimodal capabilities:
+
+**Language Model Features:**
+- **YaRN RoPE Scaling**: Advanced rope scaling for extended context lengths (up to 256K tokens)
+- **Grouped Query Attention (GQA)**: Memory-efficient attention mechanism with 8 query groups
+- **SwiGLU Activation**: Gated linear units with SiLU activation for improved performance
+- **RMSNorm**: Layer normalization without mean centering for faster computation
+- **Llama 4 Attention Scaling**: Position-dependent attention scaling for improved long-context handling
+
+**Vision-Language Features:**
+- **Vision Encoder**: Pre-trained vision encoder for robust visual understanding
+- **Multimodal Projector**: Projects vision features to language model space
+- **Flexible Image Handling**: Supports variable resolution images and multiple images per conversation
+
+## Workspace Configuration
+
+All scripts use a `WORKSPACE` environment variable to define the base directory for checkpoints and results. By default, this is set to `/workspace`. You can override it:
+
+```bash
+export WORKSPACE=/your/custom/path
+```
+
+Directory structure:
+- `${WORKSPACE}/models/` - Converted checkpoints
+- `${WORKSPACE}/results/` - Training outputs and experiment results
+
+## Checkpoint Conversion
+
+### Import HF → Megatron
+To import the HF VL model to your desired Megatron path:
+```bash
+python examples/conversion/convert_checkpoints.py import \
+  --hf-model mistralai/Ministral-3-3B-Instruct-2512-BF16 \
+  --megatron-path ${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16
+```
+
+### Export Megatron → HF
+```bash
+python examples/conversion/convert_checkpoints.py export \
+  --hf-model mistralai/Ministral-3-3B-Instruct-2512-BF16 \
+  --megatron-path ${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16/iter_0000000 \
+  --hf-path ${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16-hf-export
+```
+
+## Inference
+
+### Run Inference on Converted Checkpoint
+
+```bash
+python examples/conversion/hf_to_megatron_generate_vlm.py \
+  --hf_model_path mistralai/Ministral-3-3B-Instruct-2512-BF16 \
+  --megatron_model_path ${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16/iter_0000000 \
+  --image_path "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/resolve/main/images/table.png" \
+  --prompt "Describe this image." \
+  --max_new_tokens 100
+```
+
+Note:
+- `--megatron_model_path` is optional. If not specified, the script will convert the model and then run forward.
+- You can also use image URLs: `--image_path="https://example.com/image.jpg"`
+
+See the [inference.sh](inference.sh) script for commands to:
+- Run inference with Hugging Face checkpoints
+- Run inference with imported Megatron checkpoints
+- Run inference with exported Hugging Face checkpoints
+
+**Expected output:**
+```
+...
+Generation step 46
+Generation step 47
+Generation step 48
+Generation step 49
+======== GENERATED TEXT OUTPUT ========
+Image: https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/resolve/main/images/table.png
+Prompt: Describe this image.
+Generated: <s><s>[SYSTEM_PROMPT]You are Ministral-3-3B-Instruct-2512, a Large Language Model (LLM) created by Mistral AI, a French startup headquartered in Paris.
+You power an AI assistant called Le Chat.
+Your knowledge base was last updated on 2023-10-01.
+The current date is {today}.
+...
+[IMG_END]Describe this image.[/INST]The image presents a comparison table of technical specifications between two NVIDIA GPUs: the **H100 SXM** and the **H100 NVL**.
+
+### **FPU Performance (Floating-Point Operations Per Second)**
+- **FP64**:
+  - H100 SXM: 34 teraFLOPS
+  - H100 NVL: 30 teraFLOPS
+- **FP64 Tensor
+=======================================
+```
+
+## Finetune Recipes
+
+- See: [bridge.recipes.ministral3](../../apidocs/bridge/bridge.recipes.ministral3.md)
+- Available recipes:
+  - `ministral3_3b_finetune_config`: Finetuning for 3B VL model with PEFT support
+  - `ministral3_8b_finetune_config`: Finetuning for 8B VL model with PEFT support
+  - `ministral3_14b_finetune_config`: Finetuning for 14B VL model with PEFT support
+
+Before training, ensure the following environment variables are set:
+1. `SAVE_DIR`: checkpoint and log saving directory
+2. `HF_TOKEN`: to download models from HF Hub (if required)
+3. `HF_HOME`: (optional) to avoid re-downloading models and datasets
+4. `WANDB_API_KEY`: (optional) to enable WandB logging
+
+### Pretrain
+
+Pretraining is not verified for this model.
+
+### Supervised Fine-Tuning (SFT)
+
+See the [sft.sh](sft.sh) script for full parameter fine-tuning with configurable model parallelisms.
+
+W&B report coming soon.
+
+### Parameter-Efficient Fine-Tuning (PEFT) with LoRA
+
+See the [peft.sh](peft.sh) script for LoRA fine-tuning with configurable tensor and pipeline parallelism.
+
+W&B report coming soon.
+
+### Recommended Configurations
+
+| Model | Mode | TP | PP | Global Batch Size | Learning Rate | Hardware |
+|-------|------|----|----|-------------------|---------------|----------|
+| Ministral 3 3B | Full SFT | 1 | 1 | 32-64 | 5e-6 | 8 GPUs |
+| Ministral 3 3B | LoRA/DoRA | 1 | 1 | 64-128 | 1e-4 | 8 GPUs |
+| Ministral 3 8B | Full SFT | 2 | 1 | 32-64 | 5e-6 | 8 GPUs |
+| Ministral 3 8B | LoRA/DoRA | 1 | 1 | 64-128 | 1e-4 | 8 GPUs |
+| Ministral 3 14B | Full SFT | 4 | 1 | 16-32 | 5e-6 | 8 GPUs |
+| Ministral 3 14B | LoRA/DoRA | 2 | 1 | 32-64 | 1e-4 | 8 GPUs |
+
+**Note:** LoRA/DoRA significantly reduces memory requirements, allowing for larger batch sizes and fewer GPUs.
+
+## Evaluation
+
+Coming soon.
+
+## Hugging Face Model Cards
+
+- Ministral 3 3B Base: https://huggingface.co/mistralai/Ministral-3-3B-Base-2512
+- Ministral 3 3B Instruct: https://huggingface.co/mistralai/Ministral-3-3B-Instruct-2512
+- Ministral 3 8B Base: https://huggingface.co/mistralai/Ministral-3-8B-Base-2512
+- Ministral 3 8B Instruct: https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512
+- Ministral 3 14B Base: https://huggingface.co/mistralai/Ministral-3-14B-Base-2512
+- Ministral 3 14B Instruct: https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512
+
+## Related Docs
+- Related LLM: [Mistral](../llm/mistral.md)
+- Recipe usage: [Recipe usage](../../recipe-usage.md)
+- Customizing the training recipe configuration: [Configuration overview](../../training/config-container-overview.md)
+- Training entry points: [Entry points](../../training/entry-points.md)
+
diff --git a/examples/models/vlm/ministral3/conversion.sh b/examples/models/vlm/ministral3/conversion.sh
new file mode 100755
index 0000000000..296af05d3c
--- /dev/null
+++ b/examples/models/vlm/ministral3/conversion.sh
@@ -0,0 +1,32 @@
+#!/usr/bin/env bash
+# Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Workspace directory for checkpoints and results
+WORKSPACE=${WORKSPACE:-/workspace}
+
+# Import HF → Megatron
+uv run python examples/conversion/convert_checkpoints.py import \
+    --hf-model mistralai/Ministral-3-3B-Instruct-2512-BF16 \
+    --megatron-path ${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16
+
+# Export Megatron → HF
+uv run python examples/conversion/convert_checkpoints.py export \
+    --hf-model mistralai/Ministral-3-3B-Instruct-2512-BF16 \
+    --megatron-path ${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16/iter_0000000 \
+    --hf-path ${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16-hf-export
+
+# Round-trip validation
+uv run python -m torch.distributed.run --nproc_per_node=8 examples/conversion/hf_megatron_roundtrip_multi_gpu.py \
+    --hf-model-id mistralai/Ministral-3-3B-Instruct-2512-BF16 --tp 2 --pp 2
diff --git a/examples/models/vlm/ministral3/inference.sh b/examples/models/vlm/ministral3/inference.sh
new file mode 100755
index 0000000000..98e20c2050
--- /dev/null
+++ b/examples/models/vlm/ministral3/inference.sh
@@ -0,0 +1,45 @@
+#!/usr/bin/env bash
+# Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Workspace directory for checkpoints and results
+WORKSPACE=${WORKSPACE:-/workspace}
+
+# Inference with Hugging Face checkpoints
+uv run python -m torch.distributed.run --nproc_per_node=4 examples/conversion/hf_to_megatron_generate_vlm.py \
+    --hf_model_path mistralai/Ministral-3-3B-Instruct-2512-BF16 \
+    --image_path "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/resolve/main/images/table.png" \
+    --prompt "Describe this image." \
+    --max_new_tokens 100 \
+    --tp 2 \
+    --pp 2
+
+# Inference with imported Megatron checkpoints
+uv run python -m torch.distributed.run --nproc_per_node=4 examples/conversion/hf_to_megatron_generate_vlm.py \
+    --hf_model_path mistralai/Ministral-3-3B-Instruct-2512-BF16 \
+    --megatron_model_path ${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16/iter_0000000 \
+    --image_path "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/resolve/main/images/table.png" \
+    --prompt "Describe this image." \
+    --max_new_tokens 100 \
+    --tp 2 \
+    --pp 2
+
+# Inference with exported HF checkpoints
+uv run python -m torch.distributed.run --nproc_per_node=4 examples/conversion/hf_to_megatron_generate_vlm.py \
+    --hf_model_path ${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16-hf-export \
+    --image_path "https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16/resolve/main/images/table.png" \
+    --prompt "Describe this image." \
+    --max_new_tokens 100 \
+    --tp 2 \
+    --pp 2
diff --git a/examples/models/vlm/ministral3/peft.sh b/examples/models/vlm/ministral3/peft.sh
new file mode 100755
index 0000000000..0fb8e1b38e
--- /dev/null
+++ b/examples/models/vlm/ministral3/peft.sh
@@ -0,0 +1,62 @@
+#!/usr/bin/env bash
+# Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Workspace directory for checkpoints and results
+WORKSPACE=${WORKSPACE:-/workspace}
+
+# Common configurations
+PRETRAINED_CHECKPOINT=${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16
+MODEL_NAME=ministral3_3b
+DATASET_NAME=cord_v2
+SEQ_LENGTH=4096
+TRAIN_ITERS=50
+GLOBAL_BATCH_SIZE=32
+MICRO_BATCH_SIZE=1
+EVAL_ITERS=10
+LR=0.0002
+MIN_LR=0.00002
+LR_WARMUP_ITERS=10
+LOG_INTERVAL=1
+WANDB_PROJECT=megatron-bridge-${DATASET_NAME}
+
+# TP/PP combinations: "TP,PP"
+PARALLELISM_CONFIGS=("2,1" "1,2")
+
+for config in "${PARALLELISM_CONFIGS[@]}"; do
+    IFS=',' read -r TP PP <<< "$config"
+    
+    echo "Running LoRA finetuning with TP=$TP, PP=$PP"
+    uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
+        --recipe ${MODEL_NAME}_finetune_config \
+        --step_func vlm_step \
+        --peft_scheme lora \
+        checkpoint.pretrained_checkpoint=$PRETRAINED_CHECKPOINT \
+        model.seq_length=$SEQ_LENGTH \
+        train.train_iters=$TRAIN_ITERS \
+        train.global_batch_size=$GLOBAL_BATCH_SIZE \
+        train.micro_batch_size=$MICRO_BATCH_SIZE \
+        train.eval_iters=$EVAL_ITERS \
+        optimizer.lr=$LR \
+        optimizer.min_lr=$MIN_LR \
+        scheduler.lr_warmup_iters=$LR_WARMUP_ITERS \
+        checkpoint.save=${WORKSPACE}/results/${MODEL_NAME}_lora_tp${TP}_pp${PP} \
+        logger.log_interval=$LOG_INTERVAL \
+        logger.wandb_project=$WANDB_PROJECT \
+        logger.wandb_exp_name=${MODEL_NAME}_${DATASET_NAME}_lora_tp${TP}_pp${PP} \
+        dataset.maker_name=make_${DATASET_NAME}_dataset \
+        dataset.seq_length=$SEQ_LENGTH \
+        model.tensor_model_parallel_size=$TP \
+        model.pipeline_model_parallel_size=$PP
+done
diff --git a/examples/models/vlm/ministral3/sft.sh b/examples/models/vlm/ministral3/sft.sh
new file mode 100755
index 0000000000..193afaf10e
--- /dev/null
+++ b/examples/models/vlm/ministral3/sft.sh
@@ -0,0 +1,61 @@
+#!/usr/bin/env bash
+# Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Workspace directory for checkpoints and results
+WORKSPACE=${WORKSPACE:-/workspace}
+
+# Common configurations
+PRETRAINED_CHECKPOINT=${WORKSPACE}/models/Ministral-3-3B-Instruct-2512-BF16
+MODEL_NAME=ministral3_3b
+DATASET_NAME=cord_v2
+SEQ_LENGTH=4096
+TRAIN_ITERS=50
+GLOBAL_BATCH_SIZE=32
+MICRO_BATCH_SIZE=1
+EVAL_ITERS=10
+LR=0.00005
+MIN_LR=0.000005
+LR_WARMUP_ITERS=10
+LOG_INTERVAL=1
+WANDB_PROJECT=megatron-bridge-${DATASET_NAME}
+
+# TP/PP combinations: "TP,PP"
+PARALLELISM_CONFIGS=("2,1" "1,2")
+
+for config in "${PARALLELISM_CONFIGS[@]}"; do
+    IFS=',' read -r TP PP <<< "$config"
+    
+    echo "Running full finetuning with TP=$TP, PP=$PP"
+    uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
+        --recipe ${MODEL_NAME}_finetune_config \
+        --step_func vlm_step \
+        checkpoint.pretrained_checkpoint=$PRETRAINED_CHECKPOINT \
+        model.seq_length=$SEQ_LENGTH \
+        train.train_iters=$TRAIN_ITERS \
+        train.global_batch_size=$GLOBAL_BATCH_SIZE \
+        train.micro_batch_size=$MICRO_BATCH_SIZE \
+        train.eval_iters=$EVAL_ITERS \
+        optimizer.lr=$LR \
+        optimizer.min_lr=$MIN_LR \
+        scheduler.lr_warmup_iters=$LR_WARMUP_ITERS \
+        checkpoint.save=${WORKSPACE}/results/${MODEL_NAME}_sft_tp${TP}_pp${PP} \
+        logger.log_interval=$LOG_INTERVAL \
+        logger.wandb_project=$WANDB_PROJECT \
+        logger.wandb_exp_name=${MODEL_NAME}_${DATASET_NAME}_sft_tp${TP}_pp${PP} \
+        dataset.maker_name=make_${DATASET_NAME}_dataset \
+        dataset.seq_length=$SEQ_LENGTH \
+        model.tensor_model_parallel_size=$TP \
+        model.pipeline_model_parallel_size=$PP
+done
diff --git a/src/megatron/bridge/models/ministral3/modeling_ministral3.py b/src/megatron/bridge/models/ministral3/modeling_ministral3.py
index e541059e7a..eafe361dd0 100644
--- a/src/megatron/bridge/models/ministral3/modeling_ministral3.py
+++ b/src/megatron/bridge/models/ministral3/modeling_ministral3.py
@@ -225,7 +225,9 @@ def forward(
 
             if pixel_values is not None:
                 # Get image features using HF's method (monkey-patched)
-                image_features = self.get_image_features(pixel_values.to(inputs_embeds.dtype), image_sizes=image_sizes)
+                image_features = self.get_image_features(
+                    pixel_values.to(inputs_embeds.dtype), image_sizes=image_sizes
+                ).pooler_output
                 image_features = torch.cat(image_features, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
 
                 # Replace image tokens in text embeddings with image features
diff --git a/src/megatron/bridge/recipes/__init__.py b/src/megatron/bridge/recipes/__init__.py
index 10618ae372..ba816ad650 100644
--- a/src/megatron/bridge/recipes/__init__.py
+++ b/src/megatron/bridge/recipes/__init__.py
@@ -24,6 +24,7 @@
 from megatron.bridge.recipes.gpt import *
 from megatron.bridge.recipes.gpt_oss import *
 from megatron.bridge.recipes.llama import *
+from megatron.bridge.recipes.ministral3 import *
 from megatron.bridge.recipes.moonlight import *
 from megatron.bridge.recipes.nemotronh import *
 from megatron.bridge.recipes.olmoe import *