Build ML models
at Go speed.
A rich machine learning and accelerated math framework natively for Go. Think Jax, TensorFlow, or PyTorch, but completely Python-free. Build JIT-compiled computation graphs and execute them at scale with an OpenXLA-accelerated backend, or deploy anywhere using the pure Go backend.
Everything you need.
Nothing you don't.
Rich layers API
Dense, Conv, Attention, BatchNorm, Dropout, Embeddings — all in one package, composable, easy to extend with custom layers.
Modular design
Swap backends, optimizers, and data pipelines independently. Use just the graph layer, or the full training stack — your choice.
XLA acceleration
Compile your model graph once to XLA. The same Go code runs on CPU, NVIDIA GPU, and Google TPU without modification. Full JIT compilation, kernel fusion, and memory optimization built in.
Learn about backends →Pure Go
One go get, no virtual environments, no Python runtime, no version conflicts. Just idiomatic Go
that your whole team already knows.
Live plotting
Built-in Gonb/Jupyter integration plots training metrics in real time. Loss curves, accuracy, custom metrics — all without leaving Go.
From MNIST to transformers.
Working notebooks, command-line tools, and models you can run right now.
Gemma 3 270M
Demonstrates ONNX-converted text generation (LLM) using the Gemma 3 model.
CIFAR-10 Demo
Convolutional network image classifier available as a Jupyter Notebook.
KaLM-Gemma3 12B
Tencent's top-ranked sentence encoder loaded with go-huggingface.
AlphaZero for Hive
AI board evaluation using a GNN. Runs directly in your browser!
Questions? Join the conversation.
Ask on the #gomlx Slack channel, open a GitHub issue, or contribute an example. The library is actively maintained and the community is growing.