---
license: mit
language:
- en
tags:
- tool-use
- function-calling
- mcp
- model-context-protocol
- benchmark
- quantization
pretty_name: QuantMCP Suite
---

# QuantMCP Suite

Task fixtures and frozen tool schemas for the
[QuantMCP benchmark](https://github.com/Happynood/quant-mcp-bench), which
measures whether quantization-driven function-calling degradation (first
shown on curated BFCL/ToolACE schemas by
[QuantCall](https://huggingface.co/datasets/happynood/quantcall-suite))
carries over to real, unmodified [Model Context Protocol](https://modelcontextprotocol.io)
server tool schemas.

Unlike QuantCall's suite, everything here is originated by this project
under MIT — there is no third-party licensed data to gate or manifest-only
reference, since every task is hand-written against a real MCP reference
server's actual tool surface.

## Contents

| Path | Description |
|------|--------------|
| `tasks/u0_tasks.yaml` | 2 smoke tasks against the in-repo toy FastMCP server |
| `tasks/u1_filesystem_tasks.yaml` + `tasks/u1_filesystem/` | 13 naturalistic tasks + fixture directory (including a minimal PNG fixture for `read_media_file`) for the official `@modelcontextprotocol/server-filesystem` reference server |
| `tasks/u2_git_tasks.yaml` + `tasks/u2_git/repo.tar.gz` | 12 naturalistic tasks + a deterministic fixture git repository (packed as a tarball — a nested `.git` can't be committed as plain content) for the official `mcp-server-git` reference server |
| `tasks/u3_sqlite_tasks.yaml` + `tasks/u3_sqlite/fixture.db` | 10 naturalistic tasks + a fixture SQLite database for a self-written FastMCP wrapper (the official sqlite reference server is not currently maintained upstream — see `docs/RUN_REAL.md` in the GitHub repo) |
| `tasks/u4_memory_tasks.yaml` + `tasks/u4_memory/memory.json` | 10 naturalistic tasks + a fixture knowledge-graph seed file for the official `@modelcontextprotocol/server-memory` reference server |
| `schemas/tool_schemas.json` | Frozen live tool schemas (name, description, JSON Schema) for all 39 tools across the four real server tiers (14 filesystem + 12 git + 4 sqlite + 9 memory), dumped via `quantmcp dump-schemas`. This is the exact corpus the Schema Complexity Index (SCI) is computed from. |

Every task declares which tool it targets (`tool:` in the YAML — analysis
metadata only, never shown to the model) and is naturalistic in its
instruction (never names the tool to call), verified against the real
server's actual response before being trusted — see the GitHub repo's
`docs/RUN_REAL.md` for the full methodology and honesty caveats, including
the one live tool (`read_file`, deprecated in favor of the schema-identical
`read_text_file`) deliberately left without a task.

## How to reproduce

```bash
git clone https://github.com/Happynood/quant-mcp-bench
cd quant-mcp-bench
uv sync --extra llama-cpp
uv run quantmcp run --config configs/qwen3-0.6b-u1-sweep/Q4_K_M.yaml
uv run quantmcp dump-schemas --tiers filesystem,git,sqlite,memory --output tool_schemas.json
```

## Links

- GitHub: https://github.com/Happynood/quant-mcp-bench
- Results dataset: https://huggingface.co/datasets/happynood/quantmcp-results
- Leaderboard (Space): https://huggingface.co/spaces/happynood/quantmcp-leaderboard
- Sibling project (curated-schema quantization benchmark): https://github.com/Happynood/quant-toolcall-bench