21 NAND Docs 中文

What a large model on chain would cost

Running a large model directly in an on-chain processor on BNB Chain costs about 20,000 BNB per token. So the model runs off chain and disputes are settled on chain.

Conclusion

Direct execution: per token

The example is DeepSeek-V4.1-Flash (about 16 billion active parameters per token), running in the on-chain RISC-V processor. The numbers below are estimates from the count of multiply-adds, not measurements.

CaseInstructions per tokenGas per tokenBlocks filledDays the whole chain is fullBNB per token
20 instructions per multiply-add (measured inner loop)3.20×10^113.985×10^145,732,00129.8519,925
40 instructions per multiply-add6.40×10^117.969×10^1411,463,07859.7039,847
80 instructions per multiply-add1.28×10^121.594×10^1522,925,232119.4079,691

Run off chain, settle on chain: one settlement

SettlementEstimated gasBNB at 0.05 gwei
Submit and confirm the result for one token150,000–300,0000.0000075–0.000015
One full dispute (39 rounds of bisection, both sides, plus the final step)about 5 million–10 million0.00025–0.00050

What already works

With a tiny model of 88 parameters (16→4→4, int4), the whole dispute flow runs end to end, all locally:

Not done yet: the on-chain dispute contract and a testnet deployment.

How it was calculated