MCP server for MLX-QUANT — ternary quantization over Model Context Protocol. mode=ternary, maybe, mergeq.
  • Python 84.7%
  • Metal 15.3%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
8bit-Wraith 8df3f28eca
Merge pull request #1 from 8b-is/aye/mcp-startup
Repair MCP startup and validate packed matrix inputs
2026-10-05 20:07:02 -04:00
tests Repair MCP startup and validate packed matrix inputs 2026-10-05 20:05:51 -04:00
.gitignore chore: add .gitignore for Metal build artifacts 2026-08-12 05:56:52 +02:00
fused_ternary_matmul.metal feat(metal): add fused BitNet b1.58 ternary matvec Metal Shading Language kernel 2026-08-12 05:50:44 +02:00
pyproject.toml Repair MCP startup and validate packed matrix inputs 2026-10-05 20:05:51 -04:00
README.md Repair MCP startup and validate packed matrix inputs 2026-10-05 20:05:51 -04:00
requirements.txt Repair MCP startup and validate packed matrix inputs 2026-10-05 20:05:51 -04:00
server.py Repair MCP startup and validate packed matrix inputs 2026-10-05 20:05:51 -04:00

mlx-quant-mcp

MCP server for ternary quantization (BitNet b1.58) over Model Context Protocol using the MCP Python SDK 2.x.

Tools

  • quantize
  • dequantize
  • matmul
  • maybe_matmul
  • mergeq_matmul
  • gather_qmm
  • info

Usage

pip install -r requirements.txt
python server.py

Requires: MLX-QUANT (pip install mlx).

The 8b-is MCP Ecosystem

MCP Server Purpose
honest-irc-mcp Quantum-proof messaging + honesty-auth
ayeos-mcp Ternary matrix inference (LINOSV seed)
mlx-quant-mcp Ternary quantization (BitNet b1.58)
bluesky-mcp AT Protocol (24 tools)

axiomquant.org · pocoo.vaked.dev

The stdio server uses MCP SDK 2.3 or later (below 3). The installed mlx-quant-mcp command and python server.py use the same synchronous entry point. Install the MLX-QUANT backend separately before running.

Startup tests: python -m unittest discover -s tests. They exercise real MCP initialization and tool discovery with a synthetic MLX module; they do not validate tensor operations or GPU results.

dequantize accepts base64 of row-major, little-endian uint32 packed weights (16 two-bit codes per word). shape is the unpacked rows,cols; scales is JSON with shape rows,cols/group_size. Malformed dimensions, byte lengths and non-finite scales are rejected before tensor execution.