Training code
Model training organised on top of PyTorch.
AI Development
PyTorch Lightning services and a checkpoint security fix
We have no published PyTorch Lightning build. Lightning 2.6.6 fixed two security issues in load_from_checkpoint, and we show the nearest Python work and what to check.
0Published builds on PyTorch Lightning
2.6.6Latest release, 10 September
2Security fixes in that release
2.6.5Previous release, 27 May
What we build
What we would take on with PyTorch Lightning
Checkpoint hygiene
Only trusted checkpoints loaded, after the load_from_checkpoint fixes.
Models behind an API
A FastAPI service in front of a trained model.
Upgrades
Moves onto 2.6.6 or later.
Architecture
Where PyTorch Lightning would sit
Training code in front of the model service
Training
PyTorch LightningWhere it would sit
PyTorchModel library
Serving
FastAPIPython 3.13
Clients
React 19 appAccount briefs and grades
Illustration. A dashed outline marks where the technology would sit. We have no published build on it.
Releases
Recent Lightning releases
The latest Lightning releases
20 May
Lightning 2.6.4
A patch release.
27 May
Lightning 2.6.5
The next patch release.
10 September
Lightning 2.6.6
Marked latest. Two security fixes for arbitrary code execution in load_from_checkpoint.
Source: PyTorch Lightning releases on GitHub, checked 9 November 2026. The list shows day and month without a year.
Builds
The closest builds we can show you
AI product, in-house
Sales OS
Python 3.13 and FastAPI with an AI engine that writes a graded brief through a Claude-compatible API.
1 creditOpens all six tabs of an account brief
Logistics, fleet operations
TruckFleet Labs
A Python 3.13 back end on FastAPI with telematics ingest and PostGIS.
38Drivers on the app by week three
Fit check
When PyTorch Lightning is the right call
Pick PyTorch Lightning when
- You train PyTorch models and want tidy code
- Several people share training code
- You run training on your own servers
Look elsewhere when
- You need a published PyTorch Lightning build to inspect, which we do not have
- A model API already answers the question, as in Sales OS
- You have no labelled data yet
Explore
Same job, other tools
Pairs with
Where we used it
Stack
What PyTorch Lightning usually sits beside
PyTorchPython
MLflow
Docker
TensorFlow
KubernetesFAQs
What buyers ask first
Which Lightning version is current?
Lightning 2.6.6, marked latest on GitHub and released on 10 September. Versions 2.6.5 and 2.6.4 came out on 27 May and 20 May.
Do you have a published PyTorch Lightning build?
No. We say so on the first call and show the nearest work, which is Python 3.13 back ends and an AI product that calls a model API.
What was fixed in 2.6.6?
The release notes list two security fixes for arbitrary code execution in load_from_checkpoint. Load only checkpoints you trust.
Who owns the code and the cloud accounts?
You do, from the first commit. Repositories and cloud accounts stay in your name.
Tell us how you train your models
Thirty minutes on your training code, your checkpoints and your Lightning version. No sales pitch.