Key takeaway
- Install PyTorch with Anaconda and learn what happened to Apache MXNet (retired 2023) and Caffe (last commit 2020), with clean conda environments for each.
You can still use Anaconda to create isolated environments for deep learning work, but the three frameworks in this guide are not equal in 2026. PyTorch is the maintained choice, and its install page generates the command for your platform. Apache MXNet is retired: the Apache Attic records the retirement as September 2023, and its resources are now read-only. Meanwhile the Caffe repository shows its latest commit on 13 February 2020 and no published releases.
The older version of this guide used third-party conda channels (pjmtdw, soumith, intel) and Python 3.6. Those commands should not be copied into new projects. This guide keeps the environment workflow that still applies and tells you what to check for each framework. For production model delivery, see our posts on MLOps and DevOps for machine learning.
Which deep learning framework should you install today?
Status was read from the project pages named in the table on 13 November 2026. Confirm each on the official page before you commit a team to a framework.
| Framework | What it is | Current status | When it fits |
|---|---|---|---|
| PyTorch | Python package with tensor computation on GPU and neural networks built on an autograd system, according to pytorch.org | Maintained; the pytorch.org selector offers pip as its package manager | New projects, research and production |
| Apache MXNet | Deep learning framework supporting symbolic and imperative programming | Retired in September 2023 and moved to the Apache Attic, completed in February 2024, according to the Attic page | Only to keep an existing model running while you migrate |
| Caffe | C++ framework with Python and MATLAB interfaces, from Berkeley AI Research | Repository not marked archived, but the latest commit on the default branch is dated 13 February 2020 and no releases are listed, per GitHub | Maintaining legacy models only; avoid for new work |
How do you install Anaconda and check the download?
Download the current installer for your operating system from the Anaconda download page rather than from any fixed archive URL, because file names change with each release. Compare the SHA-256 checksum of the file with the hash published for that exact installer, where one is published, before you run it.
sha256sum Anaconda3-<version>-Linux-x86_64.sh
bash Anaconda3-<version>-Linux-x86_64.sh
source ~/.bashrc
conda list
According to the Anaconda download page, use of its offerings at an organisation of more than 200 employees or contractors requires a paid business licence unless it qualifies for discounted or free use. Miniconda is offered there as a lighter installer. Confirm the current terms before you roll either out across a company.
How do you create a clean environment for each framework?
Give every framework its own environment so that Python and CUDA-related dependencies never collide. Pick a Python version that the framework release you want lists as supported; the old Python 3.6 pin is no longer a safe default.
conda create --name pytorch_env python=<supported version>
conda activate pytorch_env
conda info --envsHow do you install PyTorch, with or without a GPU?
Open the selector on pytorch.org, choose your operating system, compute platform (CPU, CUDA or ROCm) and package manager, and run the command it prints inside the activated environment. At the time of reading, the selector listed pip, LibTorch (C++ only) and source, and did not list conda, so a pip command inside your conda environment is the route pytorch.org documents. The old soumith and anaconda channels are out of date. A conda route also exists: the conda-forge page shows conda install conda-forge::pytorch, but it does not say how to choose a CPU or GPU build, so check the build against your CUDA driver first. After installing, run Python and check the framework and GPU:
python -c "import torch; print(torch.__version__, torch.cuda.is_available())"
nvidia-smiWhat should you do about MXNet and Caffe installs?
The earlier commands for MXNet (-c pjmtdw mxnet, mxnet-cudnn) and Caffe (-c intel caffe, caffe-gpu with protobuf 3.2) relied on community or vendor channels that targeted Python 3.6. They are left out here on purpose. If you must run a legacy model, pin the environment, record the exact package versions, scan the dependencies, and plan the move to a maintained framework. Export the model to an open format where the framework supports it, and re-train or convert on PyTorch.
What does a safe adoption sequence look like?
| Step | Action | Why |
|---|---|---|
| 1 | Inventory models and the framework each one runs on | Shows where retired or inactive projects sit |
| 2 | Build one environment per framework, with versions recorded | Repeatable builds and easy rollback |
| 3 | Check status on the official project page | Avoids new work on retired software |
| 4 | Migrate legacy models to a maintained framework | Security fixes and driver support continue |
| 5 | Govern the result: owners, approvals and decision traces | Keeps models auditable once agents use them |
Once models feed AI agents, governance matters as much as installation; see our guide to agentic workflows. ElixirData Context OS keeps governed context, versioned policy and decision traces in one decision layer above AI agents, and ElixirClaw is an agentic OS platform to build, orchestrate and govern AI agents and automation workflows while keeping humans in control.
To map one model-selection decision your agents should make, and the context and policy it depends on, book a Context Workshop. See what happens in the workshop.
Frequently asked questions
Is Apache MXNet still maintained?
No. The Apache Attic records MXNet as retired in September 2023, with the move to the Attic completed in February 2024. Its resources are read-only.
Should I install Caffe for a new project?
Not advisable. The BVLC GitHub repository shows its latest commit on 13 February 2020 and no releases. For new work, a maintained framework such as PyTorch is the safer choice.
What is the right way to install PyTorch in Anaconda?
Create and activate a dedicated environment, then run the command generated by the selector on pytorch.org for your platform and compute option. The conda-forge channel is the alternative conda route.
Why use a separate conda environment per framework?
Each framework pins its own Python and library versions, so separate environments prevent conflicts and make rollbacks simple.
How do I check that PyTorch can see my GPU?
Run torch.cuda.is_available() in Python and nvidia-smi in the shell.
Follow XenonStack on LinkedIn: XenonStack on LinkedIn for new guides on agentic AI, data and platform engineering.
Explore the XenonStack platform family: ElixirData · ElixirClaw · ElixirHub · NexaStack · MetaSecure · Neural AI · AgentGRC · AgentSRE · AgentAnalyst · AgentInstruct · AgentRAI · AgentScrum · AgentSearch · AgentForce · AI Aviator · SAIF Aviator · CohortIQ · XenonStack AI · XenonStack UAE