Made for ML Engineers, by ML Engineers

Why we built
FrameTrain

We wanted a tool ourselves that makes local fine-tuning as easy as a single click – no cloud subscription, no privacy compromises, no hours of setup. So we built it.

The Story

The start was frustrating: we wanted to fine-tune a domain-specific language model on our own dataset. Sensitive data – nothing for the cloud. So: local training. The alternative was writing our own Python scripts, debugging dependencies, maintaining config files.

For someone with an ML background: doable. For everyone else: a wall of complexity. Yet the actual problem – "I want to adapt this model to my data" – is actually simple and clear.

FrameTrain is an attempt to make the gap between idea and first training run as small as possible. A native desktop app that wraps the entire stack – HuggingFace, PyTorch, LoRA, GPU scheduling – behind an interface you understand immediately.

And because we want as many people as possible to be able to use it: starting at €4.99/month. Cancel anytime, no hidden fees, no surprises.

Mission

Empower every ML engineer and researcher to train AI models locally – independent of cloud infrastructure, an internet connection, or budget. Local training should become as natural as local development.

Vision

A world where proprietary company data no longer has to land on someone else's servers just to use AI. Where a doctor, a lawyer, or a researcher can adapt a model to sensitive data – fully locally, fully securely.

Our Principles

These convictions guide every decision we make at FrameTrain.

Privacy first

Local training isn't a feature – it's the foundation. Your data belongs to you. We have no API that ever gets to see your training data.

Simplicity over complexity

ML training shouldn't fail at setup. FrameTrain aims to get your first training run started in under 10 minutes – no shell knowledge required.

Accessibility

The €4.99/month Early Access price is not an accident. We want students, researchers, and indie developers to have the same tools as big companies.

Open source where possible

We believe in open ecosystems. FrameTrain builds on HuggingFace, PyTorch, and the open-source ML stack – and gives back where we can.

Tech Stack

FrameTrain is built entirely on open-source technologies. No proprietary ML framework, no vendor-lock-in libraries. That means: when you use FrameTrain, real PyTorch runs on your machine – not some wrapper.

PyTorch
Deep learning framework
ML
HuggingFace Transformers
Model library
ML
PEFT / LoRA
Parameter-efficient fine-tuning
ML
BitsAndBytes
QLoRA quantization
ML
Tauri
Cross-platform desktop framework
App
Rust
App system backend
App
Next.js
Web interface
Web
CUDA / MPS
GPU acceleration
HW

Milestones

From the first version to the roadmap

Nov 2024

Early Access Launch

First public version with LoRA, HuggingFace integration, and GPU monitoring. 100+ early users on day one.

Dec 2024

QLoRA & BF16

QLoRA support enables training 13B models on 12 GB VRAM. BF16 for Ampere GPUs.

Jan 2025

Apple M4 Support

Full Metal Performance Shaders optimization for Apple Silicon M4 chips.

2025Planned

Roadmap

Multi-GPU training, vision models, extension ecosystem, collaborative projects.

Open Source

The website and parts of FrameTrain's infrastructure are open source. We believe transparency builds trust – especially for a tool that runs locally on your hardware.

View on GitHub
100%
Local & offline
3
Operating systems
10+
GPU architectures
€4.99
Per month in Early Access
Join the community

Ready to train locally?

Starting at €4.99/month. Cancel anytime. No cloud. Full control over your models and data.