Vaayu: Small Language Model (SLM) for Machine-to-Machine Tool Calling

Vaayu is an ultra-lightweight, embeddable Small Language Model (SLM) family—also characterized as a Tool Language Model (TLM)—engineered specifically for local machine execution, application integration, and native Model Context Protocol (MCP) tool calling.

Note: Model Context Protocol (MCP) is an open specification designed by Anthropic.

Key Features

• Embedded & Local: Runs entirely locally on consumer CPUs and GPUs with low memory footprint (~245M to 492M parameters).
• Native MCP Support: Direct, first-class connection to Model Context Protocol servers over standard I/O (stdio) and Server-Sent Events (SSE).
• Structured Tool Calling: Emits precise tool calls in structured JSON formats with built-in schema compliance.
• Low Latency: Sub-50ms Time-To-First-Token (TTFT) on modern consumer CPUs with Grouped Query Attention (4:1 GQA).
• Self-Contained: Clean Python package on top of PyTorch—zero complex orchestration frameworks or proxy daemons required.

Installation

pip install vaayu

Quickstart

1. Load Pretrained Model from Hugging Face

You can load official weights directly from the Hugging Face Hub:

from vaayu import Vaayu

# Load official Vaayu-Base directly from Hugging Face Hub
ai = Vaayu.from_pretrained("meetmendapara/Vaayu-Base")

# Generate response
response = ai.chat("Explain the purpose of Model Context Protocol (MCP) in one concise sentence.")
print(response)

2. Connect to Local MCP Tools

Vaayu natively discovers and invokes tools provided by MCP servers:

from vaayu import Vaayu

ai = Vaayu.from_pretrained("meetmendapara/Vaayu-Base")

# Connect to any local MCP server (e.g. filesystem or custom service)
ai.attach_mcp_server(
    command="npx",
    args=["-y", "@modelcontextprotocol/server-filesystem", "./workspace"]
)

# Run an agentic prompt with tool execution
result = ai.agent_step("Read the file config.json and list all defined settings.")
print(result)

3. Load Local Checkpoints

If you have trained or downloaded local weights:

from vaayu import Vaayu

ai = Vaayu.load_local("checkpoints/vaayu_base/vaayu_final.pt", variant="base")
print(ai.chat("Hello, Vaayu!"))

Command Line Interface (CLI)

Vaayu comes with an interactive CLI for chatting, testing tools, and inspecting weights:

# Start an interactive chat session with Hugging Face weights
vaayu chat --repo meetmendapara/Vaayu-Base

# Or chat with a local checkpoint
vaayu chat --weights checkpoints/vaayu_base/vaayu_final.pt

Empirical Benchmarks

Evaluated on the Vaayu MCP Benchmark Suite (200 curated scenarios across 4 core dimensions on an 8-core CPU):

Architecture Variants

Limitations

• Context Window: Vaayu-Base has a 2,048-token context ceiling. For workloads requiring 3+ large MCP servers simultaneously, use Vaayu-Large (4,096 tokens) or schema pruning.
• Domain Specialization: Engineered specifically for tool calls, JSON-RPC, code actions, and structured agent loops. Not intended for creative prose or open-domain trivia.
• Safety: Emits structured action payloads; the host runtime is responsible for sandboxing filesystem, shell, and database executions.

Resources & Links

• Hugging Face Model: meetmendapara/Vaayu-Base
• Source Code & Documentation: GitHub: Meetmendapara09/Vaayu-SLMM
• Model Context Protocol (MCP): modelcontextprotocol.io

License

MIT License. Free for academic research and commercial applications.