Vaayu (SLMM) - Small Language Machine Model

Vaayu is a high-performance Small Language Machine Model (SLMM) family engineered from scratch specifically for local machine execution, seamless application embedding, and native Model Context Protocol (MCP) tool calling.

• Vaayu-Base (245M): Live on Hugging Face 👉 meetmendapara/Vaayu-Base
• Vaayu-Large (492M): Advanced multi-tool orchestration and self-correction variant.

Key Highlights

• From Scratch Architecture: Not a fine-tune. Built with modern LLM innovations:
  • Rotary Position Embeddings (RoPE) with context window of 2,048 tokens
  • Grouped Query Attention (GQA) (16 Q heads, 4 KV heads) for 4x KV-cache memory reduction during local inference
  • SwiGLU Gated Feed-Forward Networks for superior parameter efficiency
  • RMSNorm pre-normalization for numerical stability
• Parameter Footprint: 245 Million parameters (~490MB in FP16 / ~245MB in 8-bit quantized), fitting comfortably inside consumer laptops, edge devices, and embedded desktop runtimes.
• Native MCP Bridge: First-class support for MCP client & server interactions over stdio and SSE, enabling direct connection to local IDEs, CLI tools, browsers, and filesystem agents.
• Trained on Kaggle GPUs: Sourced from premier tool-use datasets via Kaggle CLI and trained on accelerated GPU kernels under a strict 12-hour compute budget.

Repository Structure

.
├── @docs/                          # Step-by-step design, training & deployment documentation
│   ├── step-0-architecture-and-specs.md
│   ├── step-1-dataset-curation.md
│   ├── step-2-tokenizer-and-special-tokens.md
│   ├── step-3-training-from-scratch.md
│   ├── step-4-mcp-connector-and-tool-calling.md
│   ├── step-5-software-embedding-guide.md
│   └── step-6-deployment-and-huggingface.md
├── src/                            # Core model & training codebase
│   ├── model/                      # Custom Transformer architecture & configuration
│   ├── tokenizer/                  # Custom BPE tokenizer & tool tokens
│   ├── data/                       # Kaggle dataset fetcher & MCP synthesizer
│   └── training/                   # AMP training engine & Kaggle runner
├── vaayu/                          # Embeddable runtime package
│   ├── core.py                     # Local inference engine with KV caching
│   ├── mcp_client.py               # Native Model Context Protocol connector
│   └── embed.py                    # High-level embedding API for software integration
├── scripts/                        # Automation scripts (Kaggle submission, HF upload)
└── examples/                       # Embedded tool-calling & MCP usage examples

Quickstart

1. Installation

pip install -r requirements.txt
pip install -e .

2. Embedding Vaayu in Any Local Software

from vaayu import Vaayu

# Initialize embedded local model
vaayu = Vaayu.load_local("checkpoints/vaayu-final.pt")

# Connect to any local MCP server (e.g., filesystem, git, or custom app)
vaayu.connect_mcp_server(command="npx", args=["-y", "@modelcontextprotocol/server-filesystem", "./"])

# Run autonomous agentic loop
response = vaayu.chat("List all files in the current workspace and summarize their purpose.")
print(response)

Documentation

Refer to the @docs directory for complete technical specifications:

• @docs/step-0-architecture-and-specs.md
• @docs/step-1-dataset-curation.md
• @docs/step-2-tokenizer-and-special-tokens.md
• @docs/step-3-training-from-scratch.md
• @docs/step-4-mcp-connector-and-tool-calling.md
• @docs/step-5-software-embedding-guide.md
• @docs/step-6-deployment-and-huggingface.md