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

Vaayu-Large is an open-source Small Language Model (SLM)—also characterized as a Tool Language Model (TLM)—with 492,727,040 parameters, trained completely from scratch for local execution, in-process embedding, and native Model Context Protocol (MCP) tool execution.

Unlike standard conversational chatbots trained on open-ended dialogue, Vaayu is engineered as a machine-centric runtime engine: trained end-to-end directly on structured JSON-RPC 2.0 schemas, dynamic tool contracts, and autonomous local agent loops.

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

Model Details

• Model Name: Vaayu-Large
• Exact Parameters: 492,727,040 (Analytical derivation in param_count.md)
• Architecture: Custom Transformer Decoder with:
  • Rotary Position Embeddings (RoPE) ($\theta = 10,000.0$)
  • Grouped Query Attention (GQA) (4:1 ratio) reducing KV-cache allocation by 75%
  • SwiGLU Gated Feed-Forward Networks with 256-byte tensor core alignment
  • Pre-RMSNorm ($\epsilon = 10^-5$)
  • Untied Embeddings (untied input embeddings and output projection) for logit fidelity
• Context Window: 4096 tokens
• Inference Latency: Sub-50ms Time-To-First-Token (TTFT) on modern consumer CPUs.

Empirical Benchmarks

1. Berkeley Function Calling Leaderboard (BFCL v3) Standardized Evaluation

Evaluated against the official BFCL v3 / Gorilla ToolBench standardized taxonomy (250 cases on consumer CPU):

*Note: BFCL reports that general-purpose small models without native tool grammar rely on prompt-based wrapping, resulting in high syntax hallucination rates and low argument precision.

2. Vaayu MCP Benchmark Suite (200 Scenarios)

Installation & Quickstart

1. Install via PyPI

pip install vaayu

2. Stream Directly from Hugging Face Hub

from vaayu import Vaayu

ai = Vaayu.from_pretrained("meetmendapara/Vaayu-Large")
response = ai.chat("Explain the purpose of Model Context Protocol (MCP) in one sentence.")
print(response)

Tool Calling & MCP Integration

Vaayu connects directly to any local MCP server over standard input/output (stdio) and Server-Sent Events (SSE):

from vaayu import Vaayu

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

# Connect to any standard MCP server (e.g. filesystem, sqlite, terminal, git)
ai.attach_mcp_server(
    command="npx",
    args=["-y", "@modelcontextprotocol/server-filesystem", "./workspace"]
)

# Run autonomous agentic tool loop
response = ai.chat("Find all configuration files and summarize their settings.")
print(response)

Autonomous Error Self-Correction

<|im_start|>user
Inspect the production configuration file at 'config/settings.yaml'.<|im_end|>
<|im_start|>assistant
<|thought_start|>I need to read the configuration file specified by the user.<|thought_end|>
<|tool_call_start|>{"name": "read_file", "arguments": {"path": "config/settings.yaml"}}<|tool_call_end|>
<|im_end|>
<|im_start|>tool
<|tool_result_start|>{"status": "error", "error_code": "ENOENT", "message": "File not found: config/settings.yaml. Did you mean 'config/settings.json'?"}<|tool_result_end|>
<|im_end|>
<|im_start|>assistant
<|thought_start|>The file was not found (ENOENT). The host environment suggests 'config/settings.json' exists. Retrying with corrected path.<|thought_end|>
<|tool_call_start|>{"name": "read_file", "arguments": {"path": "config/settings.json"}}<|tool_call_end|>
<|im_end|>

Limitations

• Context Ceiling: 2,048 tokens on Base (4,096 tokens on Large). Connecting multiple verbose MCP servers requires schema minification.
• Domain Specialization: Specialized for structured tool contracts, JSON-RPC, and code actions; not intended for open-domain creative writing or trivia.
• Safety: Tool calls are emitted as structured payloads; the host runtime must sandbox destructive actions before execution.

Citation & Repository

@software{vaayu2026slm,
  title={Vaayu: Small Language Model (SLM) for Machine-to-Machine Tool Calling with Native Model Context Protocol (MCP)},
  author={Mendapara, Meet},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/meetmendapara/Vaayu-Large},
}
@misc{meet_mendapara_2026,
	author       = { Meet Mendapara },
	title        = { Vaayu-Large (Revision ac591dd) },
	year         = 2026,
	url          = { https://huggingface.co/meetmendapara/Vaayu-Large },
	doi          = { 10.57967/hf/10507 },
	publisher    = { Hugging Face }
}

For full architecture specifications, parameter math, training scripts, and benchmarks, visit the GitHub Repository.