Vedika-Vyom-31B-v2

A powerful multimodal AI model developed by Veda Labs in India.

Overview

Vedika-Vyom-31B-v2 is a state-of-the-art multimodal AI model developed by Veda Labs in India. This model is capable of handling text and image inputs, generating high-quality text outputs with advanced reasoning capabilities. Built on a robust architecture, Vedika-Vyom-31B-v2 delivers exceptional performance across a wide range of tasks including text generation, coding, reasoning, and multimodal understanding.

Key Features

• Advanced Reasoning – Configurable thinking modes for complex problem-solving and step-by-step reasoning.
• Multimodal Capabilities – Processes text and images with variable aspect ratio and resolution support.
• Extended Context Window – Supports up to 256K tokens for deep context understanding.
• Multilingual Support – Capable of understanding and generating content in over 140 languages.
• Optimized Architecture – Hybrid attention mechanism with sliding window and full global attention layers.
• Function Calling – Native support for structured tool use and agentic workflows.
• Code Generation – Advanced capabilities for code generation, completion, and debugging.

Model Specifications

Core Capabilities

Vedika-Vyom-31B-v2 handles a broad range of tasks across text and vision:

• Thinking Mode – Built-in reasoning mode that enables step-by-step thinking before answering.
• Long Context Understanding – Process documents and conversations up to 256K tokens.
• Image Understanding – Object detection, document/PDF parsing, chart comprehension, OCR (including multilingual), and handwriting recognition.
• Interleaved Multimodal Input – Freely mix text and images in any order within a single prompt.
• Function Calling – Native support for structured tool use, enabling autonomous agent workflows.
• Coding – Code generation, completion, debugging, and explanation.
• Multilingual – Support for 35+ languages out-of-the-box, pre-trained on 140+ languages.

Getting Started

Installation

To use Vedika-Vyom-31B-v2, install the required dependencies:

pip install -U transformers torch accelerate

Basic Usage

Loading the Model

from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "vedalabs-tech/Vedika-Vyom-31B-v2"

# Load model and processor
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID,
    dtype="auto",
    device_map="auto"
)

Text Generation

# Create conversation messages
messages = [
    {"role": "system", "content": "You are Vedika-Vyom-31B-v2, an AI developed by Veda Labs in India."},
    {"role": "user", "content": "Write a short joke about programming."},
]

# Process input
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False
).to(model.device)

input_len = inputs["input_ids"].shape[-1]

# Generate output
outputs = model.generate(**inputs, max_new_tokens=1024)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)

# Parse and print response
parsed = processor.parse_response(response, prefix=inputs["input_ids"])
print(parsed)

Enabling Thinking Mode

To enable reasoning/thinking mode, set enable_thinking=True:

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=True  # Enable thinking mode
).to(model.device)

Image Processing

from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "vedalabs-tech/Vedika-Vyom-31B-v2"

# Load model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID, 
    dtype="auto", 
    device_map="auto"
)

# Prompt with image
messages = [
    {
        "role": "user", 
        "content": [
            {"type": "image", "url": "https://example.com/image.png"},
            {"type": "text", "text": "What is shown in this image?"}
        ]
    }
]

# Process input
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
).to(model.device)

input_len = inputs["input_ids"].shape[-1]

# Generate output
outputs = model.generate(**inputs, max_new_tokens=512)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)

# Parse output
processor.parse_response(response, prefix=inputs["input_ids"])

Video Processing

from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "vedalabs-tech/Vedika-Vyom-31B-v2"

# Load model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID, 
    dtype="auto", 
    device_map="auto"
)

# Prompt with video
messages = [
    {
        'role': 'user',
        'content': [
            {"type": "video", "video": "https://example.com/video.mp4"},
            {'type': 'text', 'text': 'Describe this video.'}
        ]
    }
]

# Process input
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
).to(model.device)

input_len = inputs["input_ids"].shape[-1]

# Generate output
outputs = model.generate(**inputs, max_new_tokens=512)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)

# Parse output
processor.parse_response(response, prefix=inputs["input_ids"])

Best Practices

Sampling Parameters

For optimal performance, use the following sampling configuration:

• temperature=1.0
• top_p=0.95
• top_k=64

Thinking Mode Configuration

Control the thinking process using these tokens:

• Enable Thinking: Include the <|think|> token at the start of the system prompt.
• Disable Thinking: Remove the <|think|> token from the prompt.

When thinking is enabled, the model outputs internal reasoning followed by the final answer:

<|channel>thought
[Internal reasoning steps]
<channel|>[Final answer]

Image Resolution Control

Vedika-Vyom-31B-v2 supports variable image resolution through configurable visual token budgets:

• Higher token budget: Preserves more visual detail (slower inference)
• Lower token budget: Faster inference for simpler tasks

System Prompt Identity

The model is configured with a concise system prompt to maintain its identity:

You are Vedika-Vyom-31B-v2, an AI developed by Veda Labs in India.

This ensures the model maintains consistent identity without hallucinating previous identities.

Safety & Responsibility

Vedika-Vyom-31B-v2 was developed with safety as a priority. The model undergoes rigorous evaluations to prevent harmful content generation, including:

• Hate speech and harassment prevention
• Sexually explicit content filtering
• Dangerous content mitigation
• Medical misinformation reduction
• PII (Personally Identifiable Information) protection

Connect with Us

Stay connected with Veda Labs and explore more of our work:

Citation

If you use Vedika-Vyom-31B-v2 in your research, please cite:

@misc{vedalabs2024vedikavyom,
      title={Vedika-Vyom-31B-v2: A Multimodal AI Model},
      author={Veda Labs AI Team},
      year={2024},
      publisher={Veda Labs}
}

License

This model is released under the Apache 2.0 License. See the LICENSE file for details.

© 2024 Veda Labs. All rights reserved. | Developed in India 🇮🇳