Keryx KRX
Proof-of-Model BlockDAG: the PoW is a seeded random walk through ML model weights in VRAM — proving ownership of the model, not the correctness of its answers. VRAM and bandwidth replace hashrate as the mining resource.
About Keryx
sources: official materials · verified 2026-10-13Keryx is a BlockDAG fork of rusty-kaspa whose work function walks randomly through a neural network's weights, with the seed derived from the block header and nonce. Every read is a network round-trip, so the model cannot be precomputed, partially copied, or hosted remotely — what matters is VRAM capacity and bandwidth, which is why ASICs lose meaning here. The protocol's honesty is unusual: it proves model OWNERSHIP and leaves inference correctness to an economic mechanism (locked reward collateral, strikes, burns, suspension), because GPU floats are not bit-identical across cards. The founder-stated end goal is on-chain AI agents paying per inference; a 6-shard DeepSeek split-model test ran 4/4 across six machines, but at 3 s/token against a 0.4-0.7 s projection — an unexplained gap between demo and product.
How it works
- Network
- BlockDAG (rusty-kaspa lineage), 10 blocks/s
- Consensus
- Proof-of-Model (PoM v3): seeded random walk through model weights in VRAM; collateral-enforced 'service' availability; bit-reproducible response verification unsolved (research phase 5)
- Mining algorithm
- Proof-of-Model — VRAM + bandwidth bound
- Launched
- 2026 (H6-H13 hardfork season Aug-Sep 2026 incl. PoM v3 exploit + relaunch)
- Hardware
- GPU
Tokenomics — token & supply
- Ticker
- KRX
- Maximum supply
- 9.905B main-phase cap (founder-stated)
- Emission
- 5.10 KRX/block at 10 blocks/s, 4-year halvings; ~25% of mined supply burned to date (fees, red blocks, escrow penalties); devfund removed in v0.5.4, replaced by a consensus-level 5% R&D allocation
- Reward now
- 5.10 KRX/block (snapshot 29 Sep 2026, founder-stated)
Reference material
cited on this page · links verified 2026-10-13Who to watch
sensors, not shills — verify before actingRisks
Emission is extremely heavy: ~1.61B KRX gross per year at the founder-stated 5.10 KRX x 10 blocks/s against ~445M net circulating — year-one dilution of multiples, structurally mining-to-sell territory until halvings bite.
Explorer market cap is computed on mined supply, ignoring the ~25% burned — headline caps overstate net value by roughly a quarter until you adjust.
Proof-of-Model proves ownership, not correctness: inference quality is enforced only economically (collateral, strikes, burns). Bit-reproducible verification is explicitly unsolved (research phase 5).
The demo-to-product gap is measured: split-model serving ran 3 s/token against a 0.4-0.7 s projection with no published explanation — the service is not sellable at demo speeds.
Founding team of ~10 built from contributors with one visible founder; a community 'integrity expert' earned the role by finding two early PoM holes — unusually transparent process for the cohort.
tokenomics
- emission
- 5.10 KRX/block at 10 blocks/s, 4-year halvings; ~25% of mined supply burned to date (fees, red blocks, escrow penalties); devfund removed in v0.5.4, replaced by a consensus-level 5% R&D allocation
- reward now
- 5.10 KRX/block (snapshot 29 Sep 2026, founder-stated)
Don't confuse with
KRX — other KRX tickers: verify chain before trusting any figure — young project, thin listings
catalysts
Not financial advice. Verify before deploying capital.