Assistant
Ask anything — it routes to another connected node and pays them in credits. Switch to "Chat with my own model" above for a free, no-login local model instead (needs one loaded on the left first).
Trust boundary
nodes return raw model output. category mapping and ground-truth
scoring run on the coordinator — a node can't grade its own work.
accuracy — on 0 scored
The corpus — one tile per job, colored by the node that ran it
0nodes
0jobs done
0jobs/min
0tok/s aggregate
0tokens out
0p50 tok/s · node
0p90 tok/s · node
0:00elapsed
this swarm—
ghost: single node at current pace—
Findings assemble live
Nodes
Event feed
Model catalog — browse & compare
MMLU figures are self-reported from each model's own card/technical
report — different labs use slightly different eval harnesses, so treat as directional, not authoritative.
—means no comparable published number was found (not zero). Context is what WebLLM actually configures at runtime for every model here — currently 4096 for all of them, a browser memory constraint, not each base model's native context length.
| model | tier | vram | context | mmlu |
|---|
Account
Sign up or log in to start building a job history and
credit balance (coming next). Participating as a compute node never requires an account.