How HoloNear works

HoloNear pools idle, attested hardware — Apple Silicon Macs first, now single-GPU NVIDIA too, with confidential multi-GPU domains next — into a private, decentralized network for the full model lifecycle: pretraining, fine-tuning, RL, and serving models too large for any one machine.

The guarantee

One axiom — no single node ever holds the whole secret — bought twice over. HoloNear hides:

• Your data and prompts from the machine's owner — via hardened in-process execution, hardware attestation, and end-to-end encryption.
• The model weights from every compute provider — via pipeline sharding: each node holds only a few transformer blocks, so the full model never materializes anywhere.

Holding both at once, on untrusted residential and cloud hardware over the open internet, is what HoloNear is.

A run is a mesh

A mesh is one coordinator + N nodes (where N = num_stages × replicas_per_stage). The coordinator schedules work and relays sealed tensors between nodes — it never sees plaintext data or weights, and it is never in the FLOP path. Nodes do the actual compute on the shards assigned to them.

A researcher drives a run with holonear train (or the Python SDK); each node streams progress back, and this dashboard shows it live — loss curves, throughput, fleet, and the contributor leaderboard.

Trust & settlement run on NEAR

The hot path — routing, scheduling, microbatch dispatch — stays off-chain in the TEE coordinator. NEAR carries only the low-frequency, high-value layer: operator identity and stake, attestation-bound coordinator keys, collective ownership of protocol-trained models, and future settlement. During Genesis, contribution accounting is a live coordinator ledger; the chain removes the need to trust the coordinator for mainnet-critical state without ever touching a gradient.

Why contribute

If you run a node, you supply compute to live training and serving jobs and build Genesis Points — based on presence, GPU-seconds, and throughput, attributed to your NEAR account, and visible on the Genesis leaderboard. No cluster required: a single Apple Silicon Mac or one GPU box is enough to join — see Contribute for the three commands that do it.