Open-source AI models for local use
XYZ is an open-source family of AI models focused on local inference, programming and general text tasks.
The main goal of the project is to create useful AI models that can run locally on consumer hardware without requiring cloud APIs.
The project focuses on:
- local AI
- code generation
- instruction following
- efficient inference
- low VRAM usage
- open-source development
XYZ currently focuses on three models:
| Model | Parameters | Target Hardware | Main Focus | Status |
|---|---|---|---|---|
| XYZ-Coder Mini | 7B | 4–6 GB VRAM | Lightweight coding | Planned |
| XYZ-Coder-1 | 118M | 4–8 GB VRAM | Compact coding | In development |
| XYZ-Coder-2 | ~1B | 6–8 GB VRAM | Local coding | Planned |
Model sizes and hardware requirements may change during development.
A lightweight version designed for older GPUs and laptops with limited VRAM.
The main goal is to provide a useful coding assistant on hardware with around 4–6 GB of VRAM.
- Code generation
- Code completion
- Basic programming questions
- Local inference
- Low memory usage
The first experimental XYZ-Coder model.
| Parameter | Value |
|---|---|
| Parameters | 118,056,960 (~118M) |
| Architecture | Decoder-only Transformer |
| Layers | 14 |
| Attention heads | 12 |
| Embedding size | 768 |
| Context length | 512 tokens |
| Vocabulary | 24,000 BPE |
| Dropout | 0.1 |
Current training direction:
| Category | Target |
|---|---|
| 🔥 Code | 65% |
| 🧠 Dialogue | 20% |
| 📚 Knowledge | 15% |
The current training dataset contains approximately 923M+ tokens.
The current model is trained from scratch using a custom PyTorch training pipeline.
The next generation of XYZ-Coder.
Target size:
~1 billion parameters
The goal of XYZ-Coder-2 is to provide significantly better:
- code generation
- instruction following
- context understanding
- programming explanations
- local coding performance
Development details will be published as the model progresses.
XYZ models are built around Transformer architectures and trained using PyTorch.
Current training pipeline includes:
- AdamW
- BF16 mixed precision
- scaled dot-product attention
- cosine learning rate
- learning-rate warmup
- gradient clipping
- automatic batch-size selection
- checkpointing
- custom BPE tokenizer
XYZ-Coder-1 is currently being developed on consumer hardware.
GPU: NVIDIA RTX 4060 Ti 16GB
CPU: Intel Core i5-12400
RAM: 32GB