aarnavrexwal05/Anexa-v0.8-Thinking

Anexa (formerly MicroGPT)

Anexa is a minimal, educational, yet production-grade implementation of a GPT (Generative Pre-trained Transformer) language model.

It began as a "from scratch" Pure Python project to understand the math behind AI, and has evolved into a high-performance PyTorch model capable of training on GPUs and generating coherent text.

🚀 Why Anexa?

In a world of closed-source giants like ChatGPT and Gemini, Anexa stands for:

• Understanding: Built line-by-line to teach you how Transformers work.
• Speed: Runs locally on your machine. No queues, no latency, no internet required.
• Privacy: Your data never leaves your laptop.
• Control: You own the weights. You own the code.

📚 Documentation

We have detailed documentation for both versions of the architecture:

• Old Architecture (Pure Python): The original educational version. Learn how backpropagation and matrix multiplication work under the hood without any libraries.
• New Architecture (PyTorch): The modern version of Anexa. Features GPU acceleration, SafeTensors persistence, and standard PyTorch layers.

⚡ Quick Start

Anexa uses PyTorch and SafeTensors for maximum performance.

1. Installation

Create a virtual environment and install dependencies (requires NVIDIA GPU for CUDA support):

# Windows
python -m venv venv
venv\Scripts\activate

# Install PyTorch (CUDA 11.8) and SafeTensors
pip install torch --index-url https://download.pytorch.org/whl/cu118
pip install safetensors packaging

2. Usage

Run the model script. It will automatically download "The Adventures of Sherlock Holmes", train the model for 5000 steps, and save it to model.safetensors.

python anexagpt.py

3. Interactive Mode

After training, the script enters an interactive mode where you can chat with the model:

Training finished...
Model saved to model.safetensors

--- Interactive Generation ---
Enter a prompt to continue (or 'quit' to exit)
Prompt: Sherlock
Response: Sherlock Holmes, who was usually very late in the mornings...

✨ New: Thinking Mode (Experimental)

Anexa now supports Reasoning Tokens (<think>...</think>), similar to DeepSeek-R1.

• How it works: The model outputs a "Chain of Thought" before the final answer.
• Visualization: In chat.py, thoughts are displayed in Gray to distinguish them from the answer.
• Note: At 0.8M parameters, the reasoning is currently structural only (it mimics the format but makes arithmetic errors).

⚔️ Anexa vs The Giants

Here is how your custom 0.8M parameter model compares to state-of-the-art LLMs. While Anexa is tiny, it beats them in Privacy and Wait Time.

Model	Parameters	Training Data	Hardware	Capability	Speed
Anexa	0.8 Million	1 Book (Sherlock)	1 Laptop GPU	Basic Completion	Instant (Local)
GPT-2 (Small)	124 Million	WebText (8M Docs)	Clusters	Paragraphs	Fast
DeepSeek-V3	671 Billion	Trillions of Tokens	H100 Clusters	Reasoning/Math	Network Lag
Kimi-k2.5	hundreds of Bills	Long-context experts	Massive Clusters	2M+ Context	Network Lag
Gemini 1.5	~Trillions	Google's Index	TPU Pods	Multimodal	Network Lag
GPT-4	~1.8 Trillions	Entire Internet	Data Centers	Advanced Reasoning	Network Lag

Detailed Comparison

• Anexa (0.8M):

  • Best For: Learning how AI works, running locally on old hardware, 100% privacy, zero latency.
  • Weakness: Can only write simple sentences. Cannot code or solve math.
  • Analogy: A bicycle. You built it, you understand it, it's free, and reliable for short trips.

• ChatGPT / DeepSeek / Kimi:

  • Best For: Coding, complex math, summarizing long documents, creative writing.

  • Weakness: Requires internet, privacy concerns (data sent to servers), can be slow (queues).

  • Analogy: A commercial jet. Powerful, fast, gets you across the world, but you don't fly it and you adhere to their schedule.

  • Analogy: A commercial jet. Powerful, fast, gets you across the world, but you don't fly it and you adhere to their schedule.

🏆 Benchmark Comparison (Hypothetical)

Here is where the giants stand (and where Anexa is, humbly):

Benchmark	Qwen3-VL-235B-A22B (Thinking)	Kimi K2.5 (Thinking)	GPT-5.2 (xhigh)	Claude 4.5 Opus (Extended Thinking)	Gemini 3 Pro (High Thinking Level)	DeepSeek V3.2 (Thinking)
Reasoning & Knowledge
HLE-Full	N/A	30.1	34.5	30.8	32.2	31.5
AIME 2025	N/A	96.1	100	92.8	95.0	97.2
GPQA-Diamond	N/A	87.6	92.4	87.0	89.1	88.5
Coding
SWE-Bench Verified	N/A	76.8	80.0	80.9	78.5	79.2
LiveCodeBench (v6)	N/A	85.0	-	82.2*	84.0	86.1
Speed (Tokens / Sec)
Generation Speed	127.9	~20	~50	~30	~45	~40

🛠️ Technology Stack

• Language: Python 3.10+
• Engine: PyTorch (CUDA)
• Persistence: SafeTensors (Zero-copy, Safe)
• Architecture: Decoder-only Transformer (GPT-2 style)
  • Multi-head Self Attention
  • Feed-Forward Networks
  • Layer Normalization
  • Residual Connections

🗺️ Roadmap

• Migrate to PyTorch
• Enable GPU Training
• Implement SafeTensors
• Create Documentation
• Upload to Hugging Face
• Train on larger dataset (TinyShakespeare)
• Implement Top-K Sampling

📄 License

MIT License. Free to use, modify, and distribute.