drag to orbit · scroll to zoom · click a neuron
write strength
+ fires · − suppressed
sweep mass
gemma 3 · 270m
Every layer, every MLP neuron, steered live. The weights download once into your browser and the model runs on your GPU. No inference server exists.
- model
- Gemma 3 270M · fp16 · 18 layers · 2048 neurons/layer
- download
- 570 MB
- runtime
- checking webgpu…
nothing is uploaded; steering happens inside the forward pass
no webgpu? it falls back to wasm (slow). odd gpu with no fp16? add
?dtype=float32 and it pulls the 1.1 GB fp32 twin instead.