Inference providers

On TensorCash, an inference provider is a miner. The run that answers a customer's question also competes for the next block. The miner is paid for the answer, and can earn the block reward for the same work [1][2].

Hardware and engines

The mining software is built on two widely used programs for running AI models, adapted so that they record a proof while they work. vLLM is used for NVIDIA GPUs, and llama.cpp for CPUs and Apple Silicon [3]. Which version of vLLM is needed depends on the GPU [2].

The guide sets no minimum amount of GPU memory. Instead, pools publish their own guidance, for example a card with at least 12 GB per instance [4].

On Linux, the software needs Docker Engine with the Compose plugin, and for NVIDIA cards also the NVIDIA Container Toolkit. On macOS, Docker Desktop is enough [2].

Setting up a miner

Setting up a miner takes four steps. The full commands are in the official guide [2].

1. 1

   Check the system

   These three commands check that Docker works, that Compose is installed and that Docker can see the GPU [2].

   docker run --rm hello-world
   docker compose version
   docker run --rm --gpus all nvidia/cuda:12.8.0-base-ubuntu22.04 nvidia-smi

2. 2

   Choose how to mine

   	Broker mode	Standalone
   What it is	The mining software connects to a provider or pool and only runs the model. The other side handles the rest.	The miner also runs its own node, a full copy of the chain, and its own verifier.
   Folder	deployments/simple-worker	deployments/docker-compose/core-miner-validation-api
   Settings	BROKER_WS_URL, PROVIDER_JWT_TOKEN	WORKER_MODE=standalone, CHAIN_NAME
   Ports	8080, no inbound ports needed	8030

   In broker mode, the address and the token come from the dashboard of the provider [2]. Most pools also offer their own mining software, which connects with your wallet address and a worker name [4] (see Mining pools).

3. 3

   Start mining

   In broker mode, copy the example settings, fill in the address and the token, and start the software [2].

   cd deployments/simple-worker
   cp .env.example .env
   # set BROKER_WS_URL and PROVIDER_JWT_TOKEN in .env
   docker compose up -d

   In standalone mode, the node, the verifier and the miner start together [2].

   WORKER_MODE=standalone CHAIN_NAME=tensor-test NODE_START_FLAGS=--desktop \
     docker compose -f deployments/docker-compose/core-miner-validation-api/docker-compose.yaml up --build

4. 4

   Check that it works

   The worker answers on /health, the logs show the connection to the broker or the start of standalone mode, and proof lines appear in the logs once the miner is working [2]. Start with CHAIN_NAME=tensor-test and switch to tensor for mainnet once everything works. Test coins are available from the faucet [5].

Which models can be used

Only models registered on the chain can be used for mining. The model used in the guide is Qwen3-8B, set with MODEL_NAME=Qwen/Qwen3-8B, which is also the model behind the Bob Labs chat [2][6]. The registered models can be checked in the wallet, in the explorer or with these commands [2].

bitcoin-cli getmodelslist
bitcoin-cli getmodelregistrationstatus <model_hash>

Each model is registered at a fixed precision, the number format in which its calculations are done, such as bf16. A miner has to run the model at that precision, because a cheaper, lower-precision setup produces different scores and will be caught by verification [7].

New models are added through registration. A sponsor locks a deposit of 5 $TSC and publishes the model's identity, version and proposed difficulty. The model is registered once supporting commits appear in at least 50 of the next 100 blocks. Miners add those commits to the blocks they mine [6].

The node

The node is TensorCash Core, started with -tensor for mainnet or -tensortest for testnet. A pruned node, which keeps only recent blocks, needs at least -prune=550 [8].

Mining can also be controlled through RPC, with commands such as startmining, setminermodel and getminingmetrics [9].

Pools

A pool combines the work of many miners and shares the block rewards between them, so income arrives in smaller, more regular amounts [4].

PPLNS means that each block reward is shared over the work miners delivered in the period before that block. Fees and payouts are set by the pools and can change, so check the pool before connecting [4].

References

1. [1]TensorCash whitepaper.tensorcash.org/whitepapers/TensorCash_Whitepaper-modern.pdf
2. [2]TensorCash blog, How to mine TensorCash: a walkthrough, 27 June 2026.tensorcash.org/blog/how-to-mine/
3. [3]Mining API whitepaper.tensorcash.org/whitepapers/MiningAPI_Whitepaper-modern.pdf
4. [4]TensorCash website, Mining pools.tensorcash.org/pools
5. [5]TensorCash docs, Faucet.tensorcash.org/docs/faucet
6. [6]TensorCash blog, One chain, many models: how TensorCash runs multiple AI models, 14 August 2026.tensorcash.org/blog/how-tensorcash-runs-multiple-models/
7. [7]TensorCash blog, Trust in action: how verification caught an fp8 miner, 23 July 2026.tensorcash.org/blog/catching-a-quantized-miner/
8. [8]TensorCash blog, How to run a node, 27 June 2026.tensorcash.org/blog/how-to-run-a-node/
9. [9]TensorCash docs, JSON-RPC, mining.tensorcash.org/docs/rpc/mining