A real transformer, laid bare.

Every part of the model, labeled.

The 3D scene is not a metaphor. Every glowing element maps to a specific tensor in Phi-3-mini's compute graph.

The 3,072 points of the residual stream are laid out by PCA of the model's layer-0 qkv_proj weights — so dims read into attention together sit near each other. Each point's brightness is the live value of that residual dim on every step.

Hover an attention head — the brightness you see is that head's output magnitude.

• 32 layer rings Each ring is one transformer block. Brightness tracks the post-attention plus post-FFN residual norm for that layer — watch the signal build as the prompt flows upward.
• 32 attention heads per layer Cyan neurons on the outer ring. Each lights up proportional to its head's output magnitude. 1,024 heads in total, all live.
• FFN slab The violet 8,192-neuron expansion. By far the largest compute budget in the model. You can see it pulse as the MLP activates.
• Residual stream (3,072 dims) The highway through the network. 3,072 points, one per dim, placed by PCA of the layer-0 qkv_proj weights so functionally related dims sit near each other. Brightness on each point is the live residual value at that dim.
• KV cache strips The growing memory of past tokens. Each strip is one position; height equals cache fill for that layer.
• LM head Final projection to 32,064 vocab logits. Softmax → next token. The live top-k distribution prints to the side panel as the model decodes.

Cross-checked against reference Phi-3.

"Strict 1:1" is a strong claim, so it has to be falsifiable. Neuropulse ships with a built-in test suite that diffs the WebGPU implementation against a reference HuggingFace fp16 Phi-3-mini on a fixed set of prompts cached as reference.json. Click the wrench icon inside the demo to run it — the actual numbers from your GPU print to your browser console.

Expect tiny deltas at the hidden-state level — that's the cost of int4, not drift. What matters is the last line: identical top-1 tokens vs fp16 Phi-3 on the test set. Re-run it on your own machine in under a minute.

And you can reach in.

Watching is the easy half. The harder, more interesting half: shift-click any of the 1,024 attention-head spheres and that head's contribution to the residual stream is zeroed before O-projection — the canonical single-head ablation from the mech-interp literature, running live on the same WebGPU context that just rendered the frame. Press Run ablated and the same prompt generates twice, side by side: once whole, once missing that one head.

Or pick a layer, run a sweep, and let it ablate every head in turn. The panel paints a strip the model itself authored — each cell colored by how much the answer moved when that head went silent. Cool tones for the heads the prompt didn't need; warm ones for the heads it leaned on. About sixty seconds buys you a layer's importance map.

This is the same primitive that found "induction heads" and traced the circuits behind in-context learning — work that, until now, lived in Python notebooks tethered to CPUs and A100s. Here it runs in a browser tab, on your laptop, in real time, on a 3.8B-parameter model. The next thirty seconds of your life can look exactly like a mech-interp paper figure.

How it's built.

Four pieces. No frameworks for the inference path, no dependency soup, no clever tricks hiding the model from you.

1. WebGPU compute & WGSL 13 pipelines, 22 buffers, 292 dispatches per token. Quantization: q4f16_1. Hand-written attention and FFN kernels.
2. MLC Phi-3-mini weights The same weights as mlc-ai/Phi-3-mini-4k-instruct-q4f16_1-MLC, fetched directly from HuggingFace and cached in the browser's Cache API.
3. Three.js scene Plain WebGLRenderer. No bloom, no particles, no decorative shaders. Every pixel pulls from a real tensor on every frame.
4. PCA layout from the model's own weights Residual points are placed by PCA of layer 0's qkv_proj.weight columns; FFN points by PCA of down_proj.weight. Dims that get read or written together end up near each other, so the geometry is shaped by the model itself, not by hand.