Vela — a Generative Media Proof-of-Inference Network
Every block is one text-to-image inference, and every node verifies it by re-running the model.
Mainnet: October 14, 2026 · 14:00 UTC
What Vela isVela is a Bitcoin Core fork in which the proof-of-work is not hashing but
inference. Every mining attempt runs a deterministic text-to-image diffusion model, and the header commits to the image that was produced.
The rest of the protocol is Bitcoin's, unchanged: transactions, scripts, the UTXO set, peer-to-peer propagation and the longest-chain rule. The unit of account is the
VELA, divisible into 10
8 vel.
Bitcoin's proof-of-work secures a ledger without a trusted party, but the work itself — repeated hashing — has no value outside the protocol. Vela asks a simple question: what if the work that secures the chain is the work people actually want? Every attempt already produces an image, so the computation spent on security is the computation delivered to users, and the block subsidy pays for it.
How it worksTwo problems have to be solved to make inference into proof-of-work, and Vela solves both without zero-knowledge proofs.
1. The work must be bound to the block. The seed of the model's initial noise is derived from the header itself, so changing the nonce changes the image: an inference cannot be reused for a second attempt or precomputed in advance. The header then commits to the result through
hashPoI, a 32-byte field that extends the 80-byte Bitcoin header to
112 bytes. A miner cannot present an image that was not produced for this block.
2. The work must be verifiable. General zero-knowledge proofs of a diffusion model are far too expensive to produce at block speed, so instead the model itself is made
deterministic: sd-turbo rewritten as an exact integer program — int8 weights, 16-bit fixed-point activations, and lookup tables for every non-linearity. A C++ node, a CUDA miner and a Python reference all produce identical bytes, so every node verifies a block by re-running the same inference, on an ordinary CPU, with no GPU and no trusted setup.
Verification costs one inference; finding a block costs, on average, many. That asymmetry is the whole point.
Every block is an imageEight consecutive testnet blocks with the prompt each miner chose. The prompt is committed in the coinbase; the image is committed in the header.
Each block carries one 4×64×64 int16 latent —
32 KiB that anyone can decode into a 512×512 image with the published VAE. You can open any block on the explorer, see its prompt and its latent, and press a button that asks a node to recompute the inference and confirm the image is the honest output of the scheduled model.
Specifications| Client | Vela Core 0.1.1 (Bitcoin Core fork) |
| Block header | 112 bytes (80-byte Bitcoin header + 32-byte hashPoI) |
| Block interval | 5 minutes |
| Difficulty | ASERT (aserti3-2d), retargeted every block, 12-hour half-life |
| Model (v2) | sd-turbo as an exact integer program (W8A16), 30.5 dB vs the float model |
| Model file | 1.19 GB, hash published in the chain parameters |
| Per block | 32 KiB latent (~3.4 GB/year); headers 112 B (~11.8 MB/year) |
| Prompt | Committed in the coinbase (OP_RETURN "POI1" + token ids) |
| Verification | ~8 s on a 10-core CPU, 12 s on 4 cores, 23 s on 2 cores — no GPU |
| Mining | ~4.7 s per image on a consumer GPU (RTX 3050) |
| Supply | 21,000,000 VELA · 1 VELA = 108 vel |
| Address prefixes | vela1 (mainnet) · tvela1 (testnet) · rvela1 (regtest) |
| Ports | mainnet 19733 P2P / 19732 RPC · testnet 39733 / 39732 |
Economics| Slow start | The subsidy ramps from 10 to 50 VELA over the first 25,920 blocks (~90 days). A new chain is most exposed while few participants have joined, so early issuance is cut by 40%. |
| Halvings | Every 210,000 blocks from a 10,368-block offset, so the first halving is at block 220,368. Total supply 20,999,979.98 VELA. |
| Developer fund | 10% of every subsidy for the first 420,768 blocks (~4 years, ~1.55M VELA), enforced by consensus: a block that does not pay it is invalid. |
| Fees | Standard Bitcoin fee market; the coinbase subsidy dominates while the network is young. |
Testnet — live nowThe public test network has been running PoI v2 from block 1:
- Over 1,000 blocks, 0 failed verifications
- Full nodes recompute every block's inference independently
- Blocks mined on a GPU are accepted by CPU-only nodes
- Prompts are chosen by miners, so consecutive blocks show different images
Explore it: explorer.velaresearch.ai — every block shows its prompt, its seed, the decode of its latent, and a "recompute" button that asks a node to verify the inference.
Run it: the node guide (build, sync, wallet, mining) is at
velaresearch.ai/docs.
RoadmapPhase 1 — mainnet (October 14, 2026). The scheduled consensus model secures the chain. Miners choose prompts; every block records one prompt and the image computed to secure it.
Phase 2 — inference as a service. The same machinery becomes a platform:
- Model registry. Models are added on chain by the project through a multi-signature registration transaction recording the model hash, version, name, download location, license and a cost weight. Only models expressed as deterministic integer programs can be registered, and registrations take effect a fixed number of blocks after confirmation so nodes can fetch the file first.
- Staking for access. Instead of paying per request, a user locks coins in a stake output bound to an API key: 100 VELA staked grants one concurrent request, N × 100 VELA grants N. Unstaking waits 1,008 blocks (~3.5 days) so the same coins cannot be reused for access.
- Requests as mining attempts. A request names a model and a prompt, references the stake and is signed with the bound key. A miner serves it by putting that prompt in the coinbase of its next attempt: the attempt's inference is the user's image, and if the attempt also satisfies the target, the image becomes the block's image. Serving a request costs the miner almost nothing beyond delivery, because every attempt needs a prompt anyway.
- Verifiable results. The result returned to the user is the latent, the 80-byte header prefix of the attempt and the miner's signature. Anyone can recompute it, so untrusted machines can serve untrusted users.
Later. Mining with any registered model, with the target scaled by the model's cost weight so no model is cheaper to mine than another; and the same registry and request mechanism extended from images to higher resolutions, multi-step sampling, video and audio.
LinksUpdates- 2026-10-09 — Mainnet scheduled for October 14, 2026, 14:00 UTC. Testnet has passed 1,000 blocks with no failed verification.
- Reserved for future updates.
Nothing in this thread is investment advice and no offer of securities is made. Mining and holding VELA carries risk; the software is experimental. Always verify: every claim about a block in this thread can be checked by opening the block on the explorer and recomputing its inference yourself.
