---
license: apache-2.0
language:
- en
base_model:
- Qwen/Qwen3-VL-4B-Instruct
tags:
- vision
- multimodal
- code-generation
- qwen-vl
- peft
- unsloth
- kiel2
---

# Kiel-2-Codex

**Kiel-2-Codex** is a specialized multimodal vision-language model fine-tuned on top of **[Qwen/Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct)**. It bridges advanced visual perception with deep code generation, layout translation, and technical instruction-following.

To accommodate different hardware setups, this repository provides both quantized **4-bit** weights and full **16-bit** weights as separate directories, alongside the raw training adapter files.

---

## Repository Structure

* **`4bit/`** — Pre-quantized 4-bit weights (ideal for lower VRAM setups like a Tesla T4 or consumer GPUs).
* **`16bit/`** — Full-precision 16-bit weights (ideal for high-VRAM training/inference).
* **Root files** — Contains the raw LoRA adapter weights (`adapter_model.safetensors`), tokenizer configuration, and chat template.

---

## Quick Start (Python Inference)

You can easily load either version directly using Hugging Face `transformers` by specifying the appropriate subfolder:

### Option A: Load the 4-bit Version (Recommended for standard GPUs)
```python
import torch
from transformers import AutoModelForVision2Seq, AutoProcessor

model_id = "kiel2/kiel-2-codex"

print("Loading Kiel-2-Codex (4-bit)...")
model = AutoModelForVision2Seq.from_pretrained(
    model_id,
    subfolder="4bit",
    torch_dtype=torch.float16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_id)
```
#### Option B: Load the 16-bit Version (Full Precision)
```Python
import torch
from transformers import AutoModelForVision2Seq, AutoProcessor

model_id = "kiel2/kiel-2-codex"

print("Loading Kiel-2-Codex (16-bit)...")
model = AutoModelForVision2Seq.from_pretrained(
    model_id,
    subfolder="16bit",
    torch_dtype=torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_id)
```
##### Training Details
Base Model: Qwen/Qwen3-VL-4B-Instruct

Fine-Tuning Framework: Unsloth / Hugging Face TRL & PEFT

Task Focus: Multimodal code generation and structural layout analysis.

###### License
This model inherits the licensing structure of its base architecture under the Apache 2.0 License.