Helion-V2.0-Thinking

---

Advanced 10.2B parameter multimodal language model with 200K context, native vision, and tool use capabilities.

Key Features

• 200K Token Context Window - Process entire books and codebases
• Native Vision Understanding - Analyze images, charts, documents, and diagrams
• Function Calling & Tool Use - Structured outputs and API integration
• Strong Reasoning - Excellent performance on math, code, and logic tasks
• Multilingual Support - 12+ languages with strong performance
• Production-Ready Safety - Comprehensive content filtering and guardrails

Quick Start

from transformers import AutoModelForCausalLM, AutoProcessor
from PIL import Image

model = AutoModelForCausalLM.from_pretrained(
    "DeepXR/Helion-V2.0-Thinking",
    torch_dtype="auto",
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("DeepXR/Helion-V2.0-Thinking")

# Text generation
prompt = "Explain quantum computing in simple terms:"
inputs = processor(text=prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0], skip_special_tokens=True))

# Image understanding
image = Image.open("photo.jpg")
inputs = processor(text="What's in this image?", images=image, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0], skip_special_tokens=True))

Benchmarks

Language Understanding

Benchmark	Helion-V2.0	Helion-V2.0-Thinking	Improvement
MMLU (5-shot)	64.2%	72.3%	+12.6%
HellaSwag (10-shot)	80.5%	84.8%	+5.3%
ARC-Challenge (25-shot)	58.3%	68.7%	+17.8%
TruthfulQA MC2	52.1%	58.4%	+12.1%
GSM8K (8-shot)	68.7%	72.1%	+4.9%
HumanEval (0-shot)	48.2%	52.8%	+9.5%

Vision & Multimodal

Benchmark	Score	Notes
VQA v2	78.9%	Visual question answering
TextVQA	72.4%	Text in images
ChartQA	76.8%	Chart understanding
DocVQA	84.3%	Document analysis
AI2D	78.2%	Scientific diagrams

Tool Use & Function Calling

Benchmark	Score
Berkeley Function Calling	89.7%
API-Bank	86.4%
JSON Schema Adherence	94.8%

Model Details

• Architecture: LLaVA (Llama-2 + SigLIP vision encoder)
• Parameters: 10.2B (text: 10.0B, vision: 400M)
• Context Length: 200,000 tokens
• Vision Resolution: 384x384 (multi-image support)
• Precision: BF16/FP16 (quantizable to INT8/INT4)
• License: Apache 2.0

Hardware Requirements

Configuration	VRAM	Performance
BF16	24GB	42 tok/s (RTX 4090)
INT8	16GB	67 tok/s (RTX 4080)
INT4	12GB	89 tok/s (RTX 4070)

Use Cases

• Conversational AI - Multi-turn dialogue with long memory
• Document Analysis - Process reports, contracts, research papers
• Code Generation - Write, debug, and explain code
• Visual Understanding - Analyze images, charts, screenshots
• Data Analysis - Interpret data and create insights
• Content Creation - Articles, stories, marketing copy
• RAG Systems - Retrieval-augmented generation
• Tool Integration - Function calling and API workflows

Installation

pip install transformers torch accelerate pillow

With Quantization

from transformers import BitsAndBytesConfig

# 8-bit (16GB VRAM)
config = BitsAndBytesConfig(load_in_8bit=True)

# 4-bit (12GB VRAM)
config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4"
)

model = AutoModelForCausalLM.from_pretrained(
    "DeepXR/Helion-V2.0-Thinking",
    quantization_config=config,
    device_map="auto"
)

Advanced Features

Function Calling

import json

tools = [{
    "name": "calculator",
    "description": "Perform calculations",
    "parameters": {"expression": {"type": "string"}}
}]

prompt = f"Available tools: {json.dumps(tools)}\n\nUser: What is 127 * 89?\nAssistant:"
inputs = processor(text=prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.2)

Long Context (200K)

# Process entire documents
with open("long_document.txt") as f:
    document = f.read()  # Up to 200K tokens

prompt = f"{document}\n\nSummarize the key points:"
inputs = processor(text=prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=1024)

Multi-Image Analysis

images = [Image.open(f"image{i}.jpg") for i in range(3)]
prompt = "Compare these images and describe the differences:"
inputs = processor(text=prompt, images=images, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)

Safety Features

Built-in safety guardrails including:

• Content filtering for harmful outputs
• PII detection and redaction
• Rate limiting capabilities
• Toxicity detection
• Appropriate refusal behavior

See safety_wrapper.py for production deployment.

Limitations

• Primarily optimized for English (good multilingual support)
• Vision works best with clear, well-lit images
• Very long contexts (150K+) require substantial VRAM
• May occasionally generate incorrect information
• Not suitable for medical/legal advice without human review

Files Included

• inference.py - Full inference script with examples
• safety_wrapper.py - Production safety wrapper
• evaluate.py - Comprehensive evaluation suite
• benchmark.py - Performance benchmarking
• QUICKSTART.md - Quick start guide
• USE_CASES.md - Detailed use case examples
• safety_config.json - Safety configuration
• requirements.txt - Dependencies
• Dockerfile - Container deployment

Citation

@misc{helion-v2-thinking-2025,
  title={Helion-V2.0-Thinking: A 10.2B Multimodal Language Model},
  author={DeepXR},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/DeepXR/Helion-V2.0-Thinking}
}

License

Apache 2.0 - See LICENSE file for details.

Acknowledgments

Built with Transformers, trained on diverse open datasets. Thanks to the open-source AI community.