AnyLearning 26 Pro — free for everyone who owns 24 Pro
AnyLearning 26 Pro is here. If you bought AnyLearning 24 Pro, it is free: the same licence key, on the same machine, with nothing to re-purchase, nothing to re-issue and nothing to ask us for. Download the new version, enter the key you already have, and it activates as 26 Pro.
There is no upgrade price and no cut-off date. A 24 Pro licence is a 26 Pro licence.
Upgrading
- Download AnyLearning 26 Pro from the download page.
- Install it. Your projects, models and annotations stay exactly where they
are — everything lives in
~/anylearning-data, and the new version reads what is already there. - Open the app and enter your licence key and the email you bought it with.
You no longer have to tell the app which product your key belongs to. It works that out from the key itself, so pasting a 24 Pro key is the whole procedure. Settings → Licence will then show AnyLearning 26 Pro, noting that it came from your 24 Pro licence.
What is new
Tabular AI and Text AI in 0.26.3
AnyLearning now works with rows as well as images. Tabular AI trains local CatBoost classification and regression models from CSV, TSV, Excel, Parquet or JSON Lines. Text AI provides classical text classification, bounded lexical and fuzzy search, and deterministic evaluation of saved model, LLM or human responses.
Large datasets are stored as compressed Parquet and scanned with DuckDB. Pages load only the requested rows and columns, profiles use an explicitly marked bounded sample, and training records when it uses a reproducible sample rather than the entire table. Public Hugging Face datasets can be inspected and partially imported from converted Parquet without executing repository code.
Four attributed, licence-cleared examples are ready to download: UCI Bank Marketing, UCI Concrete Strength, Banking77 and TruthfulQA. Start with the Tabular AI guide or Text AI guide.
Auto-labeling also now defaults to the bundled SAM 2 Hiera-Small model. Its RGB preprocessing is corrected, and the guide includes a recaptured prompt, refinement and save workflow.
Keypoint detection in 0.26.2
AnyLearning can now label and train named landmarks: body joints, corners, tooling reference points, animal anatomy, or any other ordered point schema. Points are grouped by instance and retain COCO's visible, occluded and absent states through import, editing, export and training.
The included RF-DETR-Keypoint-Preview model trains the whole path and exports to ONNX before the finished model is registered. The same release also adds RF-DETR variants for object detection and instance segmentation.
Start with the keypoint detection guide and its downloadable, generated example dataset.
Instance segmentation
Mask R-CNN — ResNet-50 and ResNet-101 — is a project type of its own now, trained like everything else and exported to ONNX at the end of the run.
It was the most-requested model type, and the reason is a real limitation of the two we already had: a box cannot separate objects that overlap, and a semantic mask cannot tell two touching objects apart. Instance segmentation gives each object its own outline, which is what counting anything stacked, packed or crowded actually requires.
Class balance
A new panel counts annotations per class, split across training, validation and test.
The split chart it sits beside answers how much data, which a project can pass while still being unlearnable: a tidy 60/20/20 split says nothing about one class having four examples and another four hundred. This one says it, and it calls out the case that is otherwise silent — a class with no training examples at all, which a model cannot learn no matter how long you train it.
A calmer window
The title bar is now part of the app rather than the operating system's: rounded corners, window controls sitting in the app's own top bar, and the labelling screen rearranged to use the room that frees up. Less chrome around the image you are actually working on.
Labelling
Shapes highlight as the pointer crosses them. It matters most where several masks overlap and nothing told you which one you were about to click.
Creating a project also opens it now, rather than leaving you to find it in the sidebar.
Everything in the box
The pretrained weights every trainer starts from now ship inside the application: the detection backbones, the classification and segmentation encoders, both Mask R-CNN checkpoints, and the hand landmark model. So does auto-labelling — MobileSAM and the two smaller SAM 2 models are installed, not downloaded.
Training in 24 needed the network the first time you used each architecture, which on an air-gapped machine meant a run that prepared your whole dataset and then stopped. Nothing fetches anything now. It is the difference between "works offline" and "works offline once you have been online".
Third-party licence notices ship with the app and are readable from Settings → About, alongside the terms of use and a plain list of which model weights are included and under which licence. The Windows installer shows the notices before installing.
Under the hood
Python 3.13 and PyTorch 2.11, built and tested on Linux, macOS and Windows — CPU and NVIDIA GPU — with every project type trained end to end in the packaged app on each platform.
Still true
Everything happens on your machine. Images, annotations, trained weights, databases: none of it is uploaded, and the app works with the network switched off, apart from the one-time licence check.
It is a larger download than 24 for that reason — the weights are inside it.
If you have not tried AnyLearning before, the download page is the place to start, and Activate your licence covers setting up a key.