Decodeine Calibrator

Decodeine Calibrator is a Linux desktop application for measuring, analyzing, selecting, pruning, and merging experts in mixture-of-experts language models. It keeps the evidence and decisions behind a GGUF transformation visible: datasets, measurements, token policies, matrix composition, affinity sources, expert allocation, and export settings remain inspectable throughout the workflow.

Decodeine Calibrator is under active development. Cookbook, measurement, plan, and report formats may change until a stable release is published.

Workflow

flowchart LR
    inputs[Dataset inputs] --> preparation[Dataset preparation]
    preparation --> measurement[MoE measurement]
    measurement --> analysis[Token and expert analysis]
    analysis --> calibration[Cookbook calibration]
    calibration --> selection[Expert selection]
    selection --> export[GGUF export]

Loading

The application provides seven workbenches:

1. Cookbook owns the portable recipe, model contract, local bindings, and export configuration.
2. Dataset Preparation produces validated, self-contained measurement input.
3. Measurement runs persistent llama-moe-measure jobs.
4. Analysis explores token rules, reconstructed contexts, routed experts, and annotations.
5. Calibration combines measurements and synthetic matrices into a cooked expert-importance matrix.
6. Selection allocates output experts and evaluates merge affinity.
7. Export reviews and executes the GGUF transformation.

Requirements

• Linux
• Python 3.11 or newer
• PyQt6
• NumPy
• the official Python mcp package
• the Python gguf package
• the Python huggingface_hub package
• a C++17 compiler with OpenMP support
• a GGUF mixture-of-experts model
• a compatible llama.cpp build containing llama-moe-measure

Accelerated export requires PyTorch 2.8 or newer with an NVIDIA CUDA or AMD ROCm runtime. CPU export does not require PyTorch. Transforming supported quantized expert tensors also requires the GGML shared library from the configured llama.cpp build.

An end-to-end export acceptance run has completed successfully. CUDA acceleration has hardware-independent validation coverage, but the NVIDIA hardware path is unverified and is not part of the release acceptance scope.

The model reader and exporter support routed-MoE Mistral4, Qwen 3.5/3.6, Granite 4.0 Hybrid, Gemma4, and MiniMax-M2-family GGUF formats. See model support for architecture identifiers, tokenizers, quantization coverage, and acceptance status.

Install and run

From the repository root:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install PyQt6 numpy gguf huggingface_hub "mcp[cli]"
python -m calibrator

Open Settings → Llama.cpp Backend and select either a llama.cpp source root with a built build/bin directory or llama-moe-measure directly. The settings page can also resolve llama-server, but this release does not launch it.

Optional assistant integration is configured under Settings → MCP. Its authenticated loopback server remains off until explicitly started.

The project does not provide a packaged distribution or locked dependency set.

Documentation

Start with the user documentation. It covers installation, the complete workbench workflow, state and privacy, and troubleshooting.

Data and privacy

Decodeine Calibrator does not include model weights, datasets, measurements, or exported models. It does not send project data to a project-operated service. Processing is local unless the user explicitly selects an external source or tool.

Measurements contain token IDs and expert observations and can reveal measured content to someone with the matching tokenizer. Prepared datasets, private state, logs, plans, reports, caches, screenshots, and exported models can also contain sensitive or identifying information. Review the state and privacy guide before sharing artifacts.

License

Decodeine Calibrator is licensed under the MIT License.