Neural network architectures, hands on
Many breakthroughs came from changing how layers connect: a shortcut here, shared attention heads there, experts instead of one feed-forward. Tessera lets you load a real architecture, from a small CNN to an LLM, change its connections and find out what happens.
Free, runs in your browser, no account. Nothing you load leaves your machine.
A small CNN, live. Drag to turn it.
Three comparisons you can open in the lab. Every number is computed from the architecture, and the catalog matches each model's published parameter count.
Classic CNNs, small MLPs, an ONNX file, or any open language model's config.json from Hugging Face: Llama, Mistral, Qwen, Gemma, DeepSeek, Phi.
A clean graph with blocks you can open, and a 3D view of the data itself: feature maps, kernel cones, neurons, token vectors and attention.
Parameters, compute, memory and KV cache, the settings that differ, and every changed block marked on both drawings.
Drag in layers, connect them, add a skip, change how often a block repeats. Shapes are checked as you go, and the text version updates with you.
Small experiments in your browser, over several seeds, with accuracy, error rate, per-class F1, confusion matrices and whether the gap beats the noise.
PyTorch modules built from your connections, and Colab notebooks that train the original and your version on a free GPU.
Add a skip connection, swap the attention, add experts. Then train both and see if it mattered.
Open the lab