ClawMelt

Overview

ClawMelt is a one-stop desktop training platform for large language models. It pairs a Flutter-based Liquid-Glass interface with a Python engine, and ships ready to download LLMs straight from Hugging Face to install, fine-tune, and deploy locally.

The goal is simple: give anyone with a PC or Mac the ability to adapt a foundation model to their own use case — without a GPU cluster, without a cloud bill, and without handing their training data to a third party.

See your training results in real time with integrated testing and reporting, plus the ability to export your data and performance matrices for collaboration with others.

What you can do with it

• Pick from a curated, first-party model catalog that ClawMelt serves natively — tuned defaults, known-good checkpoints, no hunting for weights.
• Or load any LLM supported by MLX-LM, or a local GGUF / safetensors checkpoint.
• Pull datasets directly from the Hugging Face Hub, or load a local JSONL file.
• Format data as Alpaca, ChatML, or a custom prompt template.
• Run LoRA, QLoRA, or DoRA adapters with sensible defaults.
• Watch training metrics update live over WebSocket — loss curves, gradient norms, sample generations, memory usage.
• Export the trained adapter, merge it back into the base, or quantize for inference.

How your data is handled

Everything happens on your own computer. ClawMelt does not upload datasets, checkpoints, prompts, or logs to DeadStick Digital or any third party. When you download a model or dataset from Hugging Face, the connection is made directly from your machine to their servers under their own terms. See our privacy policy for the complete treatment.

Platform

Shipping: macOS 14 Sonoma or later, Apple Silicon (M1 / M2 / M3 / M4).
Planned: Windows 11 with CUDA or ROCm.

Release channel

ClawMelt is distributed as a signed and notarized Mac application. A Mac App Store build is under review; until that is live, stable releases are hosted on our website and updated through the in-app updater.