Kiel-3.2-Omni

Kiel-3.2-Omni is an advanced, fully merged 4-pillar omni-modal foundational model built upon the Janus architecture and fine-tuned by KielTech. It is optimized for deep reasoning, mathematical problem-solving, code generation, visual perception, image generation, and audio speech routing, featuring explicit chain-of-thought <thought> blocks for transparent reasoning.

---

🚀 Model Architecture & Highlights

• Base Architecture: Janus-based MultiModalityCausalLM combined with VLChatProcessor.
• Deployment Status: Fully merged standalone weights (Zero LoRA/adapter overhead required at inference).
• The 4-Pillar Omni-Modal Stack:
  • 📝 Text & Reasoning: Advanced language modeling paired with explicit <thought> chain-of-thought blocks for step-by-step auditability, math, and code synthesis.
  • 👁️ Vision (Understanding): High-density visual perception, Visual Question Answering (VQA), OCR, and chart/graph analysis.
  • 🎨 Image Generation: Cross-modal generation capabilities to synthesize images and creative assets directly from prompt pipelines.
  • 🔊 Audio (Speech Synthesis): Integrated speech generation and audio output routing workflows.

---

🛠️ Quickstart Guide

You can load and run Kiel-3.2-Omni directly using the Hugging Face transformers library and the Janus repository utilities.

1. Installation

Ensure you have the required dependencies installed:

pip install -U torch transformers accelerate safetensors
git clone [https://github.com/deepseek-ai/Janus.git](https://github.com/deepseek-ai/Janus.git)
pip install -e ./Janus

2. Loading the Model

import torch
from PIL import Image
from janus.models import MultiModalityCausalLM, VLChatProcessor

model_id = "kiel2/kiel-3.2-Omni"

print("📥 Loading processor and model...")
processor = VLChatProcessor.from_pretrained(model_id, trust_remote_code=True)
tokenizer = processor.tokenizer

model = MultiModalityCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    trust_remote_code=True
).to("cuda")

print("✨ Kiel-3.2-Omni loaded successfully and ready for inference!")

3. Example Inference

# Prepare your multimodal prompt with explicit reasoning request
conversation = [
    {
        "role": "<|User|>",
        "content": "<image_placeholder>\nAnalyze this chart, explain the underlying mathematical trend, and wrap your reasoning in <thought> tags.",
        "images": ["path_to_image.png"]
    },
    {
        "role": "<|Assistant|>",
        "content": ""
    }
]

# Process inputs
pil_image = Image.open("path_to_image.png")
prepare_inputs = processor(
    conversations=conversation,
    images=[pil_image],
    force_batchify=True
).to(model.device, dtype=torch.float16)

# Generate response
inputs_embeds = model.prepare_inputs_embeds(**prepare_inputs)
outputs = model.language_model.generate(
    inputs_embeds=inputs_embeds,
    attention_mask=prepare_inputs.attention_mask,
    pad_token_id=tokenizer.eos_token_id,
    bos_token_id=tokenizer.bos_token_id,
    eos_token_id=tokenizer.eos_token_id,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.7,
)

response = tokenizer.decode(outputs[0], skip_special_tokens=False)
print(response)

---

📜 Citation & Acknowledgments

Developed by KielTech (2026). Built as an advanced iterative fine-tune of Kiel-3-Omni (kiel2/kiel-3-Omni) upon the DeepSeek-AI Janus multimodal framework. ```