The PyTorch Certified Associate is the first official certification for PyTorch, run jointly by the PyTorch Foundation and Linux Foundation Education. It appeared on the Linux Foundation catalogue in spring 2026 and was formally announced in June. Earlier conference slides from the PyTorch Foundation had a more advanced PyTorch Certified Advanced Developer exam on the roadmap, but at the time of writing only the associate level is available. The exam runs for 120 minutes, is proctored online and is entirely multiple choice. The Linux Foundation does not publish the number of questions or the pass mark.
The blueprint has four domains: PyTorch Fundamentals at 38% (tensors, devices, training and testing loops, using models), Performance and Optimization at 26% (precision, execution optimisation, measuring performance, distributed training), Model Development at 20% (the nn building blocks) and Data Handling at 16% (datasets, DataLoaders and transforms). There are no prerequisites. The certification lasts two years, and you get one free retake and 12 months to use your voucher. Pricing is $250 for the exam alone. The Certification Pathway, added in the autumn, bundles four self-paced modules covering the same four domains with the exam for $299, which makes it the obvious choice unless you already know exactly what you are doing.
There is also a $495 bundle with the THRIVE-ONE subscription, which only makes sense if you planned to buy that subscription anyway. What I like is the weighting. A quarter of the exam on performance is unusual for an entry-level credential, and it reflects how PyTorch is used in practice: getting a model to train is easy, getting it to train fast and fit on the hardware you have is the actual job. Most beginner material stops at writing a training loop. This pushes candidates into autocast, profiling and the basics of DistributedDataParallel, and that is a good thing. What I do not like is the format.
The Linux Foundation already runs some of the best hands-on exams in the industry with the CKA and CKAD, where you solve real tasks in a terminal. PyTorch is a library you learn by writing code, and a multiple-choice exam on it is always going to reward people who read the docs over people who have debugged a shape mismatch at 2am. I suspect cost and scale drove the decision, and I understand it, but it caps how much the badge can mean. The two-year expiry also feels like a revenue decision more than a technical one. The core tensor and nn APIs have been stable for years. Compared with the alternatives, this is the only PyTorch credential with any official standing.
Google retired its TensorFlow Developer Certificate in 2024, and the cloud ML certifications from AWS, Google and Microsoft are mostly about their managed services. For someone who wants to prove framework skills without picking a cloud, the PTCA fills a real gap. For the learning itself, the free official PyTorch tutorials and Daniel Bourke's course still teach more than any exam prep will.