FAQ

It makes AI inference usable as the work that secures a blockchain, by letting a miner run an open model to answer a question. The same computation produces the answer, a proof of which model produced it, and a chance at the next block [6].

Yes. Mainnet is live, the source code is public at git.tensorcash.org, and the current state of the chain can be followed on the explorer.

Through platforms built on the network. AskTilly offers chat, agents and an API for developers, and Bob Labs sells chat with Qwen3-8B for $TSC, at 1 $TSC per million output tokens at launch [7].

21,184,153 $TSC. Emission is front-loaded to bootstrap the network and declines with every epoch [6].

Only through mining. There was no pre-mine, no token sale and no investment round [6].

It pays miners for securing the network, and it unlocks the features of the chain, such as sending payments and assets, issuing assets, registering models, trading compute derivatives and paying for verified inference [6][7].

Not for customers who pay in regular currency. Platforms such as AskTilly sell inference in regular currency and settle with miners in $TSC, so the amount of $TSC per request adjusts to its price.

No. The inference that answers a question is the same computation that competes for a block, so the model only runs once. Recording the proof adds a small cost, measured at about 3 to 12 percent of generation speed in tests with Qwen3-8B and Qwen3-32B on an NVIDIA L40S [1]. That cost depends on the size of the model's vocabulary, not on its number of parameters. A larger model with a similar vocabulary therefore loses much less speed to the proof.

The block reward in $TSC for every block they find, and income from selling the inference they produce, either directly or through a platform [6].

No. Without customer traffic, the mining software keeps the GPU busy with prompts it generates itself, so the chain keeps running. Serving customers adds payment for the answer on top of the block reward [9].

Any open model that passes the admission rules. Registering a model requires a deposit and support from miners, and each model gets its own difficulty based on the compute it needs per token, so a cheaper model does not offer an easier route to a block. Closed models cannot be used for mining, because verifiers need the weights to check the work [2][3].

Yes. Miners can use their usual optimised software, as long as the model and its precision match what was declared. Running a lower precision than declared is detected, as happened on the live network in July 2026 [4].

In tiers. Quick checks first test whether a proof is consistent with itself. A set of statistical tests follows, and finally the work is replayed with the promised model on other hardware. The verdict is measured against the normal variation that honest hardware produces [3].

No. The verifier already knows the words, so it can process them in a single parallel pass instead of generating them one by one, which costs much less than the original run. Quick checks filter out most problems first, and the full replay is only needed to validate a block, not for every proof submitted [3].

The same model gives slightly different numbers on different hardware, so exact replay would require giving up a large part of a GPU's speed. Zero-knowledge proofs of full language model inference are still too expensive at the scale of real serving. TensorCash checks instead whether the recorded numbers fall within the variation of honest hardware, which gives a statistical guarantee rather than a proof of every single calculation [3][1].

A block that breaks the rules of the chain is rejected. A block with a RED verdict counts for zero work, so it is unlikely to stay in the chain [4]. A model can also be challenged with a deposit. If the challenge gains enough support from miners, the model is banned and its registration deposit is burned [3].

Yes. A block with a RED verdict counts for zero work, and every node follows the chain with the most valid work behind it.

Yes. The verifier is open source and can be run for your own use or as a service for others [5].

No. TensorCash has no validators. Verifiers do not stake or vote. They check the work behind each block and a block that fails counts for zero work [5].

No. The chain does not require trusted execution environments. How such hardware is attested differs per platform and operator, and it has proven exploitable in practice. Platforms on the network can still offer an attested, confidential tier to customers who want it.

Post-quantum addresses have been available since block zero. They use ML-DSA, a signature standard published by NIST [8].

TensorCash was created under the pseudonym Imosuke Takakuni. The protocol asks to be judged on its code and on how it runs, both of which are public. Commercial platforms built on the network, such as AskTilly, are run by identified teams.

Like Bitcoin, the protocol should earn its standing through its code and its operation, not through the reputation of the people behind it.