Qwen2.5-VL processes images and videos at their native resolutions without forced resizing or padding. This preserves fine details and improves accuracy, especially for:

• High-resolution images with small objects
• Documents with dense text
• Videos with rapid motion or small details

Unlike traditional models that resize all inputs to a fixed size (like 224×224), Qwen2.5-VL adapts to each image's dimensions.

M-RoPE extends standard positional encoding to handle spatial and temporal dimensions. This helps the model understand:

• Spatial relationships — Where objects are located in images
• Temporal dynamics — How scenes change across video frames
• Text alignment — How visual elements correspond to text descriptions

This makes Qwen2.5-VL particularly effective for video understanding and complex visual reasoning tasks.

Qwen2.5-VL supports up to 128K tokens in its context window, enabling:

• Analysis of long-form videos
• Processing of multi-page documents
• Complex conversations about multiple images
• Detailed system prompts with extensive examples

This extended context is especially valuable for document understanding and video analysis tasks.

The Vision Transformer (ViT) encoder uses window attention to accelerate training and inference. This optimization:

• Reduces computational complexity for large images
• Enables faster training iterations
• Improves inference speed without sacrificing accuracy

NVILA-Lite uses a unique two-stage processing method:

1. Scale: Process images and videos at high spatial and temporal resolutions to capture fine details
2. Compress: Reduce visual tokens efficiently without losing critical information

This approach enables the model to handle high-resolution images and long videos while maintaining efficiency. Benefits include:

• Faster inference compared to similarly-sized models
• Lower memory requirements during training
• Efficient processing of high-resolution inputs (up to 4K images)

NVILA-Lite is optimized for the complete model lifecycle:

• Training: Reduced training costs compared to standard VLMs of similar capability
• Fine-tuning: Efficient adaptation to custom datasets with fewer compute resources
• Inference: Low latency during deployment, suitable for real-time applications
• Edge Deployment: Small enough to run on edge devices with limited resources

Despite its compact 2B parameter size, NVILA-Lite matches or exceeds the accuracy of many larger open-source and proprietary VLMs. This makes it ideal for:

• Production deployments with strict latency requirements
• Applications where model size is constrained (mobile, edge devices)
• Scenarios requiring high throughput (processing many images per second)
• Cost-sensitive deployments where compute resources are limited

Cosmos-Reason1 is optimized for tasks requiring deep understanding and logical inference:

• Visual Reasoning: Understanding cause-and-effect relationships in images
• Logical Inference: Drawing conclusions from visual evidence combined with textual context
• Multi-Step Analysis: Breaking down complex problems into logical steps
• Contextual Understanding: Considering broader context when analyzing visual information

Example use cases:

• Analyzing diagnostic images with clinical context
• Understanding complex diagrams and technical schematics
• Solving visual reasoning puzzles and problems
• Identifying anomalies that require contextual knowledge

Cosmos-Reason1 balances computational efficiency with reasoning capability:

• Optimized attention mechanisms for faster inference
• Efficient memory utilization during complex reasoning tasks
• Suitable for both cloud and edge deployments
• Supports batch processing for high-throughput scenarios

InternVL3.5 excels at processing complex visual scenes with high accuracy:

• Fine-Grained Recognition: Identifying small objects and subtle details
• Scene Understanding: Comprehending relationships between multiple objects
• Spatial Awareness: Understanding object positions and spatial relationships
• Visual Attributes: Recognizing colors, textures, sizes, and other properties

The model is particularly effective for:

• Detailed image analysis and description
• Complex scene understanding
• Fine-grained object classification
• Visual attribute recognition

InternVL3.5 effectively combines visual and textual information:

• Natural language understanding of visual content
• Generation of detailed, accurate image descriptions
• Question answering about image content
• Following complex multi-modal instructions

With 8B parameters, InternVL3.5 offers:

• Strong performance across diverse tasks
• Reasonable resource requirements for most deployments
• Good balance between accuracy and computational cost
• Suitable for both research and production use

Recommended: Qwen2.5-VL (7B) or InternVL3.5 (8B)

These models provide strong performance across diverse multimodal tasks including:

• Visual question answering
• Phrase grounding
• Image understanding and description
• Multi-image reasoning

Start with Qwen2.5-VL 7B for the best balance of performance and efficiency.

Recommended: NVILA-Lite (2B) or Qwen2.5-VL (3B)

Choose these models when:

• Deploying to edge devices with limited compute
• Requiring fast inference with low latency
• Operating under strict memory constraints
• Processing high volumes with limited resources

NVILA-Lite offers the best efficiency, while Qwen2.5-VL 3B provides broader capabilities.

Recommended: Cosmos-Reason1 (7B) or Qwen2.5-VL (32B)

These models excel at:

• Multi-step logical reasoning
• Cause-and-effect analysis
• Contextual understanding
• Complex problem-solving

Cosmos-Reason1 is optimized specifically for reasoning tasks, while Qwen2.5-VL 32B offers maximum capability across all task types.

Recommended: Qwen2.5-VL (7B or 32B)

Qwen2.5-VL can handle OCR tasks along with other multimodal capabilities:

• Document understanding
• Text extraction from images
• Multilingual text recognition
• General vision-language tasks

Coming Soon: DeepSeek OCR will be available for specialized OCR tasks and document understanding.

Recommended: Qwen2.5-VL (32B)

Choose the largest model when:

• Accuracy is the top priority
• Sufficient compute resources are available
• Deploying to production with strict quality requirements
• Handling complex, mission-critical tasks

The 32B model provides the highest accuracy across all task types but requires more computational resources.

No, the model architecture is fixed when you create a workflow. To try a different architecture, create a new workflow with the desired model. You can maintain multiple workflows with different architectures to compare performance.

Larger models require more training time and compute resources:

• 2B-3B models: Fastest training, lowest cost
• 7B-8B models: Moderate training time, reasonable cost
• 32B models: Longer training time, higher cost but best accuracy

Training time also depends on your dataset size and training settings. Check your resource usage to monitor compute consumption.

Start with Qwen2.5-VL 7B. It offers:

• Excellent performance across diverse tasks
• Reasonable training time and resource requirements
• Strong community support and documentation
• Good balance for learning and experimentation

Follow the quickstart guide to train your first model.

Yes! You can create multiple workflows within the same training project, each using a different model architecture. This allows you to:

• Compare performance across architectures
• Choose the best model for your specific use case
• Optimize for different deployment scenarios (cloud vs. edge)

Each workflow maintains its own configuration and training runs.

All currently available models support common VLM tasks like visual question answering and phrase grounding:

• Cosmos-Reason1 is optimized specifically for reasoning tasks
• General-purpose models (Qwen2.5-VL, InternVL3.5, NVILA-Lite) handle all standard VLM tasks

Coming Soon: DeepSeek OCR will be specialized for text extraction and document understanding tasks.

Choose based on your primary use case. See choosing the right model for detailed guidance.

LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning technique. NVILA-Lite uses a different training approach and doesn't support LoRA-based fine-tuning.

This doesn't limit NVILA-Lite's capabilities—it still supports full fine-tuning on your datasets. If you specifically need LoRA-based training (for memory efficiency or other reasons), choose Qwen2.5-VL or InternVL3.5 instead.