Athenea-4B-VL-Thinking

<h1 align="center">Athenea-4B-VL-Thinking</h1>

Athenea-4B-VL-Thinking is a fine-tuned version of huihui-ai/Huihui-Qwen3-VL-4B-Thinking-abliterated, specialized in multimodal reasoning, scientific problem-solving, and visual analysis.

Trained with high-quality data that combines visual content with explicit reasoning traces using <think> and </think> tags, the model is designed to perform detailed step-by-step reasoning on vision-language tasks, complex scientific problems, competitive programming, geometry, and diagram analysis.

⚠️ Important Note: This model uses an abliterated (uncensored) base version, providing total expressive freedom and unrestricted output generation. Users are fully responsible for any use or content produced by the model. It is intended exclusively for research and experimentation purposes.

🎯 Model Description

Athenea-4B-VL-Thinking extends the structured reasoning capabilities of Huihui-Qwen3-VL toward scientific and multimodal domains, showing outstanding performance in logical problem-solving, visual analysis, and understanding complex diagrams.

Key features:

• Step-by-step visual reasoning within <think> blocks
• Specialization in scientific and analytical tasks (Chemistry, Physics, Geometry, Graph Algorithms)
• Uncensored output generation for complete reasoning visibility
• Enhanced logical consistency through focused fine-tuning
• Compatible with open inference frameworks (Transformers, vLLM, etc.)

The model was fine-tuned using the Aquiles-ai/Athenea-VL dataset, which includes 20,913 high-quality examples with diverse visual content, structured reasoning chains, and natural language explanations across multiple scientific domains.

Note: Fine-tuning was performed using Kronos, Aquiles-ai's proprietary enterprise fine-tuning system.

💻 Usage

Installation

uv pip install transformers torch accelerate qwen-vl-utils

Basic Inference

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
import torch

model_id = "Aquiles-ai/Athenea-4B-VL-Thinking"

model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id,
    attn_implementation="flash_attention_2",  # Requires flash-attn
    device_map="auto",
    trust_remote_code=True,
    dtype=torch.bfloat16,
)

# Without flash-attn:
# model = Qwen3VLForConditionalGeneration.from_pretrained(
#     model_id,
#     device_map="auto",
#     trust_remote_code=True,
#     dtype="auto",
# )

processor = AutoProcessor.from_pretrained(model_id)

image_path = "multimodal_problem.jpg"

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image", "image": f"{image_path}",
            },
            {"type": "text", "text": "A mass $m_1$ of 9.1 kg is positioned on a frictionless plane inclined at an angle of $50°$. It is tethered by a rope that passes over a frictionless pulley to a second, hanging mass $m_2$ of 7.8 kg, as depicted in the diagram below. Your task is to calculate the acceleration of this two-mass system and the tension within the connecting rope."},
        ],
    }
]

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

# Inference: Output generation
generated_ids = model.generate(**inputs, max_new_tokens=40960)
generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

Production Deployment with vLLM

Start server:

vllm serve Aquiles-ai/Athenea-4B-VL-Thinking \
  --host 0.0.0.0 \
  --port 8000 \
  --api-key dummyapikey \
  --mm-encoder-tp-mode data \
  --limit-mm-per-prompt '{"image":2,"video":0}' \
  --chat-template chat_template.jinja \
  --max-model-len=16384 \
  --gpu-memory-utilization=0.90 \
  --reasoning-parser qwen3

Request to server from OpenAI client:

from openai import OpenAI
import base64

client = OpenAI(base_url="http://0.0.0.0:8000/v1", api_key="dummyapikey")

def encode_image(image_path):
    with open(image_path, "rb") as image_file:
        return base64.b64encode(image_file.read()).decode("utf-8")

image_base64 = encode_image("multimodal_problem.jpg")

response = client.chat.completions.create(
    model="Aquiles-ai/Athenea-4B-VL-Thinking",
    messages=[
        {"role": "system", "content": "IMPORTANT: Always wrap your thinking process between <think> and </think> tags."},
        {"role": "user",
        "content": [
            {"type": "text", "text": "What's in this image?"},
            {
                "type": "image_url",
                "image_url": {
                    "url": f"data:image/jpeg;base64,{image_base64}",
                },
            },
        ],
    }],
    max_tokens=2048,
    extra_body={
        "add_generation_prompt": True,
        "enable_thinking": True,
    },
    stream=True
)

for chunk in response:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

vLLM Benefits: 20-30x faster inference, OpenAI-compatible API, continuous batching, asynchronous scheduling, support for multiple images.

🔬 Model Capabilities

Athenea-4B-VL-Thinking excels at:

• Scientific reasoning: Physics, chemistry, and mathematics problems with diagrams
• Competitive programming: Analysis of visual data structures and algorithms
• Advanced geometry: Interpretation of complex geometric figures
• Graph algorithms: Understanding and analysis of graph representations
• General multimodal analysis: Combining visual and textual information

📊 Training Dataset

The model was trained on the Aquiles-ai/Athenea-VL dataset, which contains:

• 20,913 high-quality examples with structured reasoning traces
• Diverse scientific domains: Chemistry, Physics, Competitive Programming, Geometry, Graph Algorithms
• Chain-of-Thought reasoning: All examples include explicit thought processes in <think> tags
• Balanced distribution: Randomly mixed data to prevent training biases

Contact

• More about Aquiles-ai.

• Aquiles-ai on GitHub.

• Our collections on HuggingFace.

Aquiles-playground

Work is still underway to make it compatible with Aquiles-playground

<p align="center"> Made with ❤️ by <a href="https://github.com/Aquiles-ai">Aquiles-ai</a> </p>