An open teaching tool for language models. Run a tiny model (135M–500M) right in your browser on WebGPU — watch it generate, or drive it through an agent state machine that turns one forward pass into a think → tool → observe loop. Every layer's activity — norms, heads, router flow, logit lens, attention, residual trajectories — streams live as it runs. OpenAI-compatible, testable, honest. The animation behind this card is the real visualization, on mock data.
Every layer's residual decoded as if the network stopped there. Dots mark where the prediction changes with depth — red where it settles on the token actually emitted. Settle depth ≈ per-token difficulty.
Each token's hidden state as a point in a fixed random 3D basis, connected in generation order — blue dotted trace is the full prompt context. Loops = repeated structure; outliers = attention sinks. Drag-rotate, per-layer minimap in the real app.
Select any token and see where it looked: top attention sources (mean over heads) drawn as arcs over the token strip, with per-token uncertainty (entropy + margin) riding on every frame.
A 135M model can't reliably use tools on its own — a harness makes it. A finite state machine drives the model through an explicit loop, and you watch every decision live in your browser:
Real tools run in the page (calculator, unit-convert, clock) alongside clearly-labelled sandbox tools. When a tiny model flubs the format, the harness catches it — the failure is shown, not hidden. No GPU? A scripted demo runs the same machine with no download.
Point any SDK or agent
framework at http://localhost:8080/v1 — zero code change
and every request is fully instrumented, recorded and replayable.
Spec-compliant
tool_calls round-trips (LFM2 pythonic + JSON parsing) and
image content parts — vision-language chats are hooked like any
other.
The curator records every run and gates CI: internals checks (finite norms, settle depth, uncertainty, attention focus) plus tool-reasoning suites — should-call / abstain / arguments / grounded chains.
Curated models on WebGPU via transformers.js — no install. Run one as an agent → through a state-machine harness, or use the observable-ONNX export that graph-surgeries residual streams into declared outputs, so serverless runs still show real per-layer activity.
| capability | server | browser | browser·observable |
|---|---|---|---|
| chat + top-k predictions | ✓ | ✓ | ✓ |
| tool calling · multimodal | ✓ | — | — |
| norm heatmaps · heads · lens · arcs | ✓ | — | resid |
| residual trajectory + context | ✓ | — | ✓ |
| curator recording | ✓ | — | — |
Works with any HF causal LM — hybrid conv/SSM, MoE and vision-language models are first-class. Apple-silicon (mps), CUDA and CPU devices.
product six-pager (PDF) · the vault (architecture & flow graphs, protocol, engines) · CONTRIBUTING · AGENTS.md keeps humans and AI agents honest about docs.
suite cases & curator checks (no ML needed) → curated browser engines → dashboard views for attention/entropy → observable-export coverage → hooks for new architectures.
Honest pixels (never render unmeasured data) · every PR keeps the vault diagrams true · suites green before merge · MIT licensed.