One API key, one base URL — every frontier model behind it. The empty space at the center is what makes the wheel useful.
“Thirty spokes share one hub. It is the empty space at the center that makes the wheel useful.”
Bittensor named its network after the Tao. We took our name from the founding text of Taoism — written 2,500 years before anyone needed a model router.
OpenAI, Anthropic, Google, DeepSeek, Qwen and more — any model on the catalog is a spoke. A conductor delegates each task to the spoke it picks, watches the outcome, and re-delegates if that spoke failed.
The conductor never writes an answer and never touches the question. The only things that ever reach a worker are the task text and a model ID — so there is nothing the hub could smuggle, and nothing it needs to be trusted with.
A gateway run by one company is just another gatekeeper. Thirty Spokes runs as a Bittensor subnet: miners compete to train the conductor, and one published verdict per window says who holds the crown and why.
Point your existing client at the hub. Ask for model="auto" and the network routes each request — or pin any spoke by name.
Miners route each request to the best spoke for quality per dollar.
Pin a specific provider model by name whenever you need to.
Streaming, tool use and structured output pass through unchanged.
A miner ships a conductor — full weights on a pinned architecture. The validator runs it against the sitting king on the same fresh slice of agentic benchmarks, paid on the miner's own model-marketplace key, and scores quality minus what that quality cost at an exchange rate measured from the pool rather than chosen. The crown moves only to a challenger that wins broadly. Until someone beats the best fixed policy, the king's share burns.
Open the dashboardfinalb = qualityb − λb · spendb / Cb final = mean over benchmarks
Build the conductor with the highest benchmark quality per dollar. You ship weights; the validator runs them. Nothing you write ever executes anywhere, and nothing your model writes ever reaches a worker — the competition is which model you pick, and what you spend.
Full weights on the pinned Qwen3.6-35B-A3B, by any route — fine-tune, distil, merge. Every window publishes the decision traces of the arms the owner ran, so the training set is free.
The validator's own admission gate, on your tree, locally: architecture field by field, tokenizer by digest, every tensor's name and shape, safetensors only, no code. Run it until it says admitted.
One submission per hotkey, ever. Upload ~70 GB to a private bucket, against a credential scoped to your own prefix, then commit the manifest digest on chain — that signal is what spends the shot.
The validator runs your conductor and the king on the same slice, same scaffold, same catalog, each paid on its own key. Quality minus what it cost, equal weight per benchmark.
Beat the king broadly — Δ over ε, lower bound over zero, leave-one-out and median both positive — and take 0.85 of emissions. Only then are your weights published. The five most recent ex-kings share 0.15.