What it is
PyTorch is the deep learning framework most AI research and open models are built on. It gives you tensors, automatic differentiation and GPU acceleration from Python, and it is governed by the PyTorch Foundation under the Linux Foundation.
The current release line is 2.14, which arrived on 2 September 2026; the patch release 2.14.1 followed on 30 September. Builds cover NVIDIA CUDA, AMD ROCm, Intel XPU and Apple silicon, and 2.14 adds Python 3.15 wheels, including the free-threaded (no-GIL) build, although torch.compile does not yet support Python 3.15.
The framework is free under a BSD-3-Clause licence. There is no paid edition; the cost is the hardware you run it on.
Our review
PyTorch is the layer the rest of open-source AI stands on, and nothing in 2026 threatens that. Transformers, TRL, Unsloth, vLLM and diffusers all build on it, and the repository has about 105,000 GitHub stars. For most teams the question is not whether to use it but how far down the stack to work.
The 2.14 release shows where the effort now goes: the compiler and the cluster. NVGEMM brings generated CUTLASS kernels into Inductor, the torch.compile backend, including NVFP4 matrix maths for Blackwell GPUs. Fault tolerance became a first-class part of the c10d distributed layer, with process groups that can be reconfigured in place, and a preview NCCL backend ported from torchcomms is meant as a drop-in replacement. These matter to people training across many GPUs; someone fine-tuning on one card will barely notice them.
The awkward parts are familiar. Matching a wheel to your CUDA or ROCm version still trips people up, and every minor release lists backwards-incompatible changes; 2.14 alters how gradients split at clamp boundaries, which can shift results without raising an error. Checkpoint loading is the security weak spot. torch.load with weights_only=True was meant to be the safe path, yet two advisories (CVE-2025-32434, rated critical, and CVE-2026-24747, rated high and fixed in 2.10.0) showed that a malicious checkpoint could still run code. Safetensors files sidestep the problem.
JAX keeps a foothold in some research labs and Apple's MLX is a lighter option on Macs, but neither has PyTorch's breadth of third-party code. For someone who only wants to fine-tune a model, PyTorch is the engine rather than the steering wheel, and a library such as Transformers or Unsloth on top gets the same result in a fraction of the code.
Where it earns its place
An ML engineer continuing pretraining of a 1B-parameter model on a rented 8x H100 node
Train across eight GPUs without writing a distributed system from scratch
FSDP and the c10d layer shard the model across GPUs, and 2.14's fault tolerance lets a process group recover instead of restarting the job. At Lambda's $3.99 per H100-hour on an 8-GPU instance the node costs about $32 an hour, so a lost run is the expensive failure to design against.
A PhD student implementing a new attention variant for a paper
Prototype an architecture and make it fast enough to train
Write it in eager PyTorch, debug it line by line, then wrap it in torch.compile for kernel fusion. On a Mac, the MPS backend in 2.14 gained native SVD, QR and Cholesky, so more of the linear algebra stays on Apple silicon instead of falling back to the CPU.
A platform lead standardising tooling for a 50-person ML team
Pick a framework that hiring, libraries and vendors all support
PyTorch costs nothing to licence under BSD-3-Clause and nearly every open model ships PyTorch-compatible weights, so the real costs are GPUs and engineers. The risk to manage is the supply chain: insist on safetensors checkpoints and keep torch at 2.10 or later.
Before you commit
- Setup
- pip install torch pulls the default wheel; for a specific CUDA, ROCm or CPU build, the selector on pytorch.org gives the matching index URL. With drivers already in place a first GPU tensor takes minutes; driver and toolkit mismatches are the usual delay.
- Integrations
- Transformers, diffusers, vLLM, Lightning and most training libraries sit on top; torch.export and ONNX cover deployment outside Python. Hardware backends include CUDA, ROCm 7.14, Intel XPU and Apple MPS, and Inductor already targets NVIDIA Rubin (sm_107).
- Data and privacy
- The library sends nothing anywhere on its own. The risk is loading untrusted pickle-based checkpoints: CVE-2026-24747 allowed code execution with weights_only=True on versions up to 2.9.1. The BSD-3-Clause licence permits commercial use without restriction.
- Cost at scale
- The software is free and there is no enterprise edition; compute is the bill. Lambda lists one H100 PCIe at $3.29 an hour and an A10 (24 GB) at $1.29, so a single always-on H100 is roughly $2,400 a month.
- Limits
- About 17,700 open issues and pull requests on GitHub. torch.compile does not support Python 3.15 in 2.14, and each minor release lists backwards-incompatible changes, so pin versions in production.
What's good
- The base of nearly every open-model library
- torch.compile and distributed training built in
- Wheels for CUDA, ROCm, Intel XPU and Apple silicon
- Neutral Linux Foundation governance
What's not
- Too low-level for routine fine-tuning
- Matching wheels to CUDA or ROCm versions is fiddly
- Minor releases change behaviour, so versions need pinning
- Pickle-based checkpoints remain a code-execution risk
Best for
Anyone building, researching or deeply customising neural networks.
Alternatives
PyTorch: common questions
- Is PyTorch worth it?
- Still the framework under almost every open model, and the 2.14 release keeps its lead in compilers and multi-GPU training. We rate it 4.9 out of 5.
- How much does PyTorch cost?
- Free and open source under the BSD-3-Clause licence. No paid edition.
- What are the downsides of PyTorch?
- Too low-level for routine fine-tuning; matching wheels to CUDA or ROCm versions is fiddly; minor releases change behaviour, so versions need pinning; pickle-based checkpoints remain a code-execution risk.
- Who is PyTorch best for?
- Anyone building, researching or deeply customising neural networks.
- How much does PyTorch cost for a team or in production?
- The software is free and there is no enterprise edition; compute is the bill. Lambda lists one H100 PCIe at $3.29 an hour and an A10 (24 GB) at $1.29, so a single always-on H100 is roughly $2,400 a month.
- Is PyTorch safe for company data?
- The library sends nothing anywhere on its own. The risk is loading untrusted pickle-based checkpoints: CVE-2026-24747 allowed code execution with weights_only=True on versions up to 2.9.1. The BSD-3-Clause licence permits commercial use without restriction.
- What are the best PyTorch alternatives?
- The closest alternatives we have reviewed are Hugging Face Transformers and MLX.
Answers checked Oct 23, 2026.
Make PyTorch? Show this rating on your site
