Proof of work for AI hardware.
ZeroFlux is a proof-of-work network on NEAR. ZFX is its token, created only by mining with AI hardware: NVIDIA and AMD data-center GPUs, Apple Silicon or any CPU. NEAR keeps the ledger and referees every block.
The world is building more compute than ever.
Billions of dollars are flowing into AI infrastructure. Data centers are expanding, GPUs are becoming critical assets, and inference demand continues to grow.
But demand isn't constant. Even the most sophisticated inference operations experience fluctuations in utilization. When expensive hardware isn't generating revenue, its costs don't disappear. ZeroFlux gives that hardware a second workload.
One GPU. More ways to earn.
Your infrastructure shouldn't depend on a single source of demand. ZeroFlux lets GPU owners mine ZFX on NEAR with the same tensor cores they use for AI. During testnet, ZFX has no monetary value.
Serve AI models and earn revenue from inference workloads.
Use available GPU capacity to perform proof-of-work computations and compete for ZFX mining rewards.
Your hardware. Your choice. Run on dedicated mining hardware or alongside an existing GPU infrastructure stack, subject to available compute resources.
Proof of work, reimagined for the AI era.
Each mining attempt runs a 64 → 4096 → 4096 → 4096 → 4096 → 256 network of integer matrix layers, about 51.6 million multiply-adds, then hashes the output. Miners batch thousands of attempts into ordinary GEMMs on int8 or FP8 tensor cores. Every value is a small integer, so int8, FP8, Metal and CPU arithmetic all produce the identical result, bit for bit.
Matrix-native
Computational work centered on matrix operations, rather than conventional hash-only mining.
Hardware-accelerated
Runs on NVIDIA and AMD tensor cores through cuBLAS and hipBLASLt, and on Apple Silicon GPUs through MLX and Metal.
Independently verifiable
Anyone can recheck a block on a CPU in milliseconds. To prove a bad block wrong, you name one value and NEAR recomputes just that value.
NEAR keeps the ledger.
ZeroFlux settles on NEAR: ZFX is a standard NEAR token, minted only for proof of work. Miners submit winning attempts to a NEAR contract with a bond. A block counts unless someone proves it wrong during a challenge window; then the reward is minted. No separate chain, no bridge.
How ZFX settles on NEARAnyone with compatible hardware can mine. Difficulty retargets on-chain every block.
Posting a block costs about 6 Tgas of NEAR gas. Proving a bad block wrong costs about 4.5 Tgas: NEAR regenerates one row of weights and recomputes one value.
A cheater's bond goes to whoever catches them, and every block built on a bad one is reverted. The watcher ships with the mining tools.
ZFX works with NEAR wallets and trades through NEAR Intents directly, with nothing to bridge.
No new class of hardware. A new opportunity for existing compute.
Your GPU is already capable.
Built for the GPUs already deployed across AI infrastructure. No mining ASIC required. Data-center GPUs mine roughly 2,500 to 4,000 times faster than a laptop CPU, by design.
NVIDIA
int8 (cuBLAS) and FP8 (cuBLASLt) paths, both exact.
≈ 86 µJ per attempt
AMD
int8 (hipBLAS) and FP8 (hipBLASLt) paths, both exact.
≈ 93 µJ per attempt
Apple Silicon
GPU mining through MLX or PyTorch MPS.
24x the same chip's CPU
Any CPU
Rust and C++ miners; watchers recheck blocks on any CPU.
| Hardware | Attempts / s | Real blocks mined |
|---|---|---|
| AMD MI355X | 13,124,061 | 57 in 120 s, all re-verified by an independent CPU node |
| NVIDIA H100 80GB | 8,078,510 | 42 in 120 s, all re-verified by an independent CPU node |
| Apple M5 Pro GPU (MLX) | 80,131 | 88 in 30 s on an easy local target, re-verified |
| Apple M5 Pro CPU (18 threads) | 3,293 | 55 in 25 s on an easy local target, re-verified |
Measured on the production proof-of-work profile. Every GPU path passed 278 bit-exactness checks against the reference and a worst-case accumulator stress test. We are also developing support for NVIDIA Blackwell (B200, B300), RTX 50 series and RTX PRO 6000, and AMD MI300X.
Make more of the infrastructure you already own.
AI infrastructure is capital-intensive. Every GPU represents an investment in compute capacity, power, cooling, and operations. Whether you operate a large inference cluster, rent GPU capacity, or own high-performance hardware, ZeroFlux gives it a second workload that runs on the same tensor cores.
No new class of hardware. A new opportunity for existing compute.
Get StartedFrom source to mined blocks in minutes.
Everything runs locally: a NEAR sandbox, the ZFX contract, a watcher and your miner. One command checks every claim on this page and writes an evidence report.
git clone https://github.com/zerofluxprotocol/zeroflux && cd zeroflux # build, then check bit-exactness against the published test vectors cargo build --release && make -C miner/cuda cpu ./target/release/zpu-cli selftest # local NEAR sandbox + ZFX contract + watcher + mining proxy make localnet & ./target/release/zpu-cli mine --node http://127.0.0.1:9561 # mine on an Apple Silicon GPU instead ./target/release/zpu-cli mine --node http://127.0.0.1:9561 \ --backend exec --exec ".venv/bin/python miner/python/zpu_miner.py --backend mlx"
make check # lint, unit tests, test vectors in Rust, C++, Python, MLX make e2e-near # mining on NEAR, cheaters caught at every layer, ZFX on NEAR Intents make security-test # fuzzing, sanitizers, audit, reproducible build make acceptance # all of the above, then ACCEPTANCE_REPORT.md
Built in the open.
Apache-2.0. The design documents, threat model and test evidence are published below, including what is not done yet: independent cryptanalysis, an external audit, and a mainnet monetary policy.
Put your compute to work.
Proof of work for the AI era, settled on NEAR.