[ANN] KERYX | Live AI Inference | Proof of Model | GPU PoW | 10 BPS

[HdCe]

With latest keryx-labs miner on windows downloading models is very slow and when downloading that 2200mb file whole miner closes/exits when approx. 1800mb of the file is downloaded.

[ocminer]

I've released a new version of the miner:

https://github.com/ocminer/keryx-miner-supr/releases/tag/v.0.5.3

works on all pools, optimized also for AMD and most bugs should be ironed out.

Pool https://krx.suprnova.cc

[HdCe]

I've released a new version of the miner:

https://github.com/ocminer/keryx-miner-supr/releases/tag/v.0.5.3

works on all pools, optimized also for AMD and most bugs should be ironed out.

Pool https://krx.suprnova.cc

I was using your miner.. just updated to 0.5.3 still same

I am getting "OPoI challenge in progress - PoW paused, stand by" and mining does not start. Sometimes mines for a while and then this message again. Also having "stratum message payload is empty". Using baikalmine pool.

What is causing this?

edit: tested suprnova.cc and it works.

[chafer99]

If anyone has problems running miner for AMD in Win, you can try - mc_miner_krx: https://github.com/xbin0r/mc-miner-krx/releases/tag/v0.39

[Keryx Labs]

Proof-of-Model (PoM)

Keryx was built for a single purpose: decentralized, uncensored AI inference. This means the network's value relies entirely on miners actually hosting the AI models. Until now, this was based on a "cooperative" model—relying on good faith. But good faith isn't security. Any pool or modified miner could simply strip the models away, collect rewards, and contribute absolutely nothing to the network's core mission.

PoM fixes this natively at the protocol level.

How it works: With PoM, the model is the proof-of-work. Mining becomes a memory-hard process that reads thousands of random chunks directly from the model weights stored in your GPU's VRAM. No weights in VRAM = no valid proof = no valid block. It is completely deterministic.

What it brings to the table:

- Pool-proof & Fork-proof: You cannot fake model possession or route around the system. Hold the model, or you don't mine.

- ASIC-resistant by design: The work is strictly bound to memory (weights) rather than raw compute power.

- Bigger model = bigger reward: Block rewards scale dynamically based on the model tier you prove you are hosting. This incentivizes miners to run heavier, more capable models instead of the lightest ones.

Honest scope:

PoM guarantees possession—proving that a miner genuinely holds a specific open model in VRAM. It does not yet verify that every inference output is correct; that is a separate layer currently under development. We'd rather ship a rock-solid guarantee today than overpromise.

The result: a network where mining and AI capability become one and the same. The hashrate is the model.

[Keryx Labs]

Holder-Reward — Keryx rewards those who hold the network

"Intelligence is the message. Those who carry it deserve to be rewarded."

We're rolling out a new economic mechanism: the Holder-Reward. The idea is simple — the more $KRX you hold, the more of your mining reward you keep.

The principle

On every block, the miner's share is scaled by a single ratio:

your $KRX balance  ÷  your production over the last 24h

In other words: how many days of your own mining output you're holding. The more you keep, the higher your rate climbs — up to 100 %. Whatever you don't hold is burned :fire:.

The brackets

HOLD (days of your output)   →   YOU KEEP
──────────────────────────────────────────
< 1 day                      →    40 %
≥ 1 day                      →    52 %
≥ 3 days                     →    64 %
≥ 1 week                     →    76 %
≥ 15 days                    →    88 %
≥ 1 month                    →   100 %

"1 day of output" = what your GPU mines in 24h. The bar scales automatically with your hashrate.

Why it's fair

The ratio is relative to YOUR production: a solo miner and a large farm follow exactly the same rule ("hold N days of your output"). No edge for the big players. Want the full rate? Hold ~1 month of what you mine.

What's in it for you

- No lock, no staking, non-custodial. Your coins stay 100 % yours, spendable anytime. The network just reads the balance of your payout address.
- Zero extra steps. No transaction, no signature — auto-binding happens on your payout address.
- Deflationary. The unkept delta is burned → $KRX scarcity rewards those who hold.

When

Activation: Friday, June 26, 2026, 18:00 UTC (DAA 37,780,000), as part of the upcoming hardfork.

Hold the network. The network holds you.
Keryx Node v1.2.7 — The Proof-of-Model Hardfork

⚠️ Mandatory upgrade — coordinated mainnet hardfork

▎ Activates at DAA 37,780,000 — Friday June 26, 2026, ~18:00 UTC.

This release changes the mining algorithm and activates three protocol upgrades at the same height. Every node must run v1.2.7, and every miner/pool must run a PoM-capable miner with the pinned models, before H — otherwise their blocks are rejected and they fork off the chain.

What activates at H

Proof-of-Model (PoM) — new mining algorithm

Replaces kHeavyHash. A valid block now carries a cryptographic proof of model possession: the miner proves it holds a specific open LLM in VRAM (Merkle-anchored weights, Fiat-Shamir witness verified after PoW, 100 % node-deterministic — the node never holds the weights).

- 1 GPU = 1 tier: Gemma-3-4B → Dolphin-Llama3-8B → Qwen3-32B → Llama-3.3-70B
- Pool-proof, fork-proof, genuinely anti-ASIC
- Scope honesty: PoM proves possession, not inference correctness (a separate, later layer).

OPoI v2 — uncensored model lineup

DAA-gated swap to the uncensored inference lineup.

💎 Holder-Reward — your miner cut, weighted by what you hold

Your share scales by your $KRX balance ÷ your own 24h production — i.e. how many days of your own output you hold. The shortfall is burned (deflationary). No lock, no staking, non-custodial, zero extra transaction.

< 1 day of your output → keep 40 %
≥ 1 day → 52 %
≥ 3 days → 64 %
≥ 1 week → 76 %
≥ 15 days → 88 %
≥ 1 month → 100 %

Tier-Reward — bigger model, bigger share

The miner cut also scales by your cryptographically proven PoM tier (6-point steps), compounding multiplicatively with Holder-Reward. Shortfall burned; total block reward unchanged.

Tier 0 — Gemma-3-4B → 82 %
Tier 1 — Dolphin-Llama3-8B → 88 %
Tier 2 — Qwen3-32B → 94 %
Tier 3 — Llama-3.3-70B → 100 %


Under the hood (since v1.2.6)

- PoM: PomProof transport over P2P (borsh) + RPC submit_block; weightless node-side verifier (blake3, seed-bound); per-tier R_T/N pinned; GPU-friendly mix64 seed/pow folds (K=256); kHeavyHash header check skipped under PoM.
- Holder-Reward (ratio-reward) full stack: per-SPK balance index kept in lockstep with the UTXO set; windowed-production index (O(1) slide + reorg-safe); consensus-safe inline bracket (no per-block store); fast-sync reconstruction from the pruning-point snapshot; full unit + E2E suite incl. reconstruction-equality.
- Consolidated hardfork activation banner.

Fixes

- sink_search VP-stall fixed (PoM-gated): reaches valid sinks hidden below disqualified blocks.
- Inference capability challenge: rotates on ai:cap change; re-challenge interval 6 min → 1 h.
- Docker: build the keryxd binary instead of kaspad — thanks @Namp88 (#10).

Upgrade instructions

- Node operators: pull v1.2.7, rebuild, restart before DAA 37,780,000.
- Miners & pools: switch to the v1.2.7 PoM miner with the pinned GGUF models loaded before H. A miner still on kHeavyHash — or without the models — has its blocks rejected after the fork.

Activation

Height (DAA): 37,780,000
Target time: Fri 2026-06-26 ~18:00 UTC

[Keryx Labs]

keryx-miner v0.3.3 — Proof-of-Model hardfork

https://github.com/Keryx-Labs/keryx-miner/releases/tag/v0.3.3-OPoI

This is the mining-algorithm hardfork release. At DAA 37,780,000 (Fri June 26, 2026 · 18:00 UTC) the network switches from kHeavyHash to Proof-of-Model (PoM) and swaps to the uncensored model lineup (OPoI v2). This is bundled with node v1.2.7 (PoM + OPoI v2 + holder-reward, all at the same height).

Action required before June 26, 18:00 UTC

- Upgrade to v0.3.3. A miner still on v0.3.2 will keep producing kHeavyHash blocks after H and fork off the network — its blocks will be rejected.
- Let the models download. The client auto-fetches the GGUF weights for your tier (pinned IPFS CIDs). Start the client well ahead of H so the download + the one-time possession index build are done before the cutover.
- Pool miners: PoM requires your pool to send the real chain DAA (keryx-stratum-v2 protocol). Confirm your pool/bridge is updated, or solo-mine.

Proof-of-Model (PoM)

- Mining no longer hashes kHeavyHash — the GPU walks over the model weights resident in VRAM, and a winning nonce ships a Merkle possession proof (byte-exact, pool- and fork-proof).
- 1 GPU = 1 tier. Bigger VRAM serves a bigger model and proves a higher tier.
- Zero-dup VRAM sharing: the possession walk shares the same resident matrices the inference engine uses — no duplicate copy. Inference takes priority and hot-swaps the model when a request comes in.
- Disk-backed Merkle index so the 70B tier fits; proof chunks are pread from the GGUF instead of a host RAM copy (saves ~1x model size, ~42 GB on the 70B).
- Clean kHeavyHash to PoM transition: lazy load, legacy lineup until H, no false stall/challenge at the cutover.

Uncensored model lineup (OPoI v2)

DAA-gated hot-swap from the legacy lineup to the abliterated lineup at H:

- Gemma-3-4B - abliterated, 6GB
- Dolphin3.0-8B - 8GB
- Qwen3-32B - abliterated, 24GB
- Llama-3.3-70B - abliterated, 48 GB

The client only announces models your GPU VRAM can actually serve (capability gate).

GPU / CUDA performance

- Ported Suprnova fork optimizations — Keccak unroll, sm_80 launch bounds, per-arch auto workload — thanks to @.ocminer
- eor3/bcax lop3 for Keccak theta/chi (bit-exact, +6.6% permutation) — thanks to @gooszillabtc
- Native PTX dispatch for sm_90 (Hopper: H100/H200/GH200), sm_80 (A100 / CMP 170HX), and sm_120.
- PTX regenerated for sm_61/75/80/86/89/90/100/120.
- AMD: v_dot8 default + gfx906 JIT fix (also from @.ocminer 's port).
- candle 0.8 -> 0.9.2 (CUDA 12.2 OK — unlocks Qwen3 / Gemma3 / Mistral archs).

Fixes & cleanup

- Mainnet DAA gates set for the v1.2.7 hardfork (PoM + OPoI v2 @ 37,780,000; SALT v2/v4 restored).
- Reset Qwen3 KV cache per inference (fixes broadcast_add shape mismatch on Qwen3-32B).
- One model per tier under PoM + coherent launch logs.
- Quieter logs: show OPoI inference pause instead of false "stalled" warnings; quiet model-prefetch at startup.
- Removed --cpu-inference — inference is GPU-only.
- Removed --vram-pool — incompatible with PoM.

[Keryx Labs]

Keryx Miner v0.3.4

https://github.com/Keryx-Labs/keryx-miner/releases/tag/v0.3.4-OPoI

New Keryx miner release — the full 5-model lineup, broader GPU support, multi-GPU performance fixes, and a big batch of community contributions.

What's new

5-model lineup
Your GPU is automatically matched to a model tier:

- Qwen3-1.7B (--very-light) - 4GB VRAM
- Gemma-3-4B (--light) - 8GB VRAM
- Dolphin-Llama3-8B (no flag/default) - 12GB VRAM
- Qwen3-32B (--high) - 24 GB) VRAM
- Llama-3.3-70B (--very-high) - 32 GB VRAM

The more powerful your GPU, the bigger the model you serve to the network.

Mining & reliability
- Multi-GPU PoM — mining is no longer serialized across GPUs (shared host index + per-GPU miners)
- Volta (SM70) support — older NVIDIA cards can now mine
- Failed inferences are never submitted or rewarded
- Post-fork the miner only pulls the uncensored lineup — no more downloading the dead legacy models
- Sparse-checkpoint Merkle tree — ~63× disk reduction + auto-clean of legacy trees
- HiveOS — self-locating scripts, ISO-timestamp log parsing, and correct per-device stats on rigs with an iGPU

Community contributions
This release is powered by our contributors — huge thanks to all of you:
@GooszillaBtc  — parallel PoM mining across GPUs (shared host index + per-GPU miners)
@Fear  — sparse-checkpoint Merkle tree: ~63× disk reduction + legacy-tree auto-clean
@Lukretsiy  — correct sparse-checkpoint Merkle build & path via fold_levels (co-authored)
@Dragonkeepr  — SM70 (Volta) PTX resource (#11) + iGPU per-device stats fix
@CLusmi  — HiveOS self-locating scripts + env_logger ISO-timestamp parser, and the iGPU off-by-one stats fix
@Dizzztroyer * — scoped GPU uninstall() to the inference device only, fixing hashrate loss on multi-GPU rigs (#12)

[Keryx Labs]

Keryx-node-v1.3.0-OPoI

https://github.com/Keryx-Labs/keryx-node/releases/tag/v1.3.0-OPoI

Consensus hardening + a fresh-node IBD fix. Node operators should upgrade before the soft-fork activation at DAA 46,500,000 (~July 9, 2026). No mining-algorithm change (PoM + lineup unchanged) — miners are unaffected.

Highlights

- Fresh-node IBD fix — trusted blocks no longer fail headers-proof sync with PoM possession proof missing (#14).
- H2 inference_reward minimums enforced — adds consensus floors for Qwen3-1.7B (0.3 KRX) and LLaMA-3.3-70B-Q2 (4.0 KRX), missing since the H2 lineup. Soft-fork, gated at DAA 46,500,000.
- Ratio-reward index bounded — the production index no longer grows unbounded on pruned nodes; entries below the reward window are collapsed into per-SPK floors during pruning. Storage-only, consensus-neutral, no gate.

⚠️ Action required

The inference_reward minimums are a soft-fork gated at DAA 46,500,000. Rebuild and restart keryxd before then. Upgrading avoids following a soon-to-be-orphaned block if an underpaying AiRequest is ever mined after activation.

Upgrade

git pull && cargo build --release
Restart with your existing datadir — no resync required.

What's Changed

- fix(ibd): preserve PoM data for trusted blocks by @Dizzztroyer in #14
- fix(consensus): enforce H2 inference_reward minimums (Qwen3-1.7B + 70B-Q2), gated at DAA 46_500_000
- feat(consensus): bound the ratio-reward prefix index — collapse pre-window entries into per-SPK floors during pruning

[Keryx Labs]

keryx-miner-v0.3.5-OPoI

https://github.com/Keryx-Labs/keryx-miner/releases/tag/v0.3.5-OPoI

Heterogeneous-rig mining lands, plus a mixed-rig VRAM fix and a leaner runtime. No consensus/protocol change — this is miner-side only.

Highlights

- Per-GPU multi-tier mining — a mixed rig now mines a different tier per GPU (each card runs the largest tier its VRAM holds, capped by your flag; smaller cards downgrade instead of failing the whole rig). Inference is routed to the GPU that already holds the model (so only that card pauses), and a GPU that OOMs its tier is auto-downgraded to a smaller downloaded model.
- Mixed-rig VRAM gate fixed — the ai:cap capability check now reads each GPU's VRAM from the CUDA driver instead of nvidia-smi's first line, which mismatches the CUDA device order on mixed rigs.
- Lower idle CPU & RAM — fewer idle async workers on many-core hosts, hashrate logging moved off the async runtime, bounded OPoI collections, and a non-blocking reconnect.

Single-GPU rigs are unaffected — behavior is identical.

Upgrade

Download the build for your OS below (HiveOS / Linux), or:

git pull && cargo build --release

What's Changed

- feat(pom): per-GPU multi-tier mining on heterogeneous rigs — thanks @Fear, @CLusmi
- fix(cuda): source the ai:cap VRAM gate from the CUDA driver (device 0), not nvidia-smi line order — thanks @CLusmi, @Fear
- perf(runtime): cap tokio async workers + move hashrate logger off the async runtime; yield on reconnect (Closes #13) — thanks @Fear
- reduce RAM usage on multi-core rigs and long OPoI sessions by @Fear in #14

[Keryx Labs]

📱 Keryx Wallet for Android is here — first public release!

The official self-custodial Keryx wallet is now live. Your keys, seed and password never leave your device — everything is signed locally on your phone.

What's inside:

- Create / restore a 24-word wallet, encrypted on-device
- Send & receive KRX — QR scanner, address book, live balance
- Fingerprint / Face unlock
- Live KRX price + Trade shortcut
- On-chain AI inference right from the app
- Consolidate for miners — one tap compounds your whole UTXO set
- Fully open-source — audit the code yourself

Download: https://github.com/Keryx-Labs/keryx-mobile-wallet/releases/tag/v1.0.0

Grab keryx-wallet-v1.0.0.apk, open it on your phone, allow install from this source. You can verify the download against the published SHA-256.

Feedback and issues welcome — this is our first mobile release and we want to hear from you! 💚

[Keryx Labs]

The new models scheduled for the upcoming hard fork are now available for download:


Direct download :
https://keryx-labs.com/EXAONE-4.0-1.2B.zip
https://keryx-labs.com/Mistral-7B-v0.3.zip
https://keryx-labs.com/GLM-4-9B-0414.zip
https://keryx-labs.com/Qwen3.6-27B.zip
https://keryx-labs.com/Kimi-Linear-48B.zip

Torrents :
https://keryx-labs.com/EXAONE-4.0-1.2B.zip.torrent
https://keryx-labs.com/Mistral-7B-v0.3.zip.torrent
https://keryx-labs.com/GLM-4-9B-0414.zip.torrent
https://keryx-labs.com/Qwen3.6-27B.zip.torrent
https://keryx-labs.com/Kimi-Linear-48B.zip.torrent

[Keryx Labs]

The storm has passed🔥

Following last week's attack and a necessary hard fork to secure the network, our systems are officially back online.

A huge thank you to our entire community for your patience and continuous support during this difficult period!

[Keryx Labs]

Keryx Miner v0.4.1 — performance & compatibility release⚡

PoM kernel optimizations
• Vectorized chunk reads (`ulonglong2`): ~+47% hashrate on GDDR6X + ILP×2: ~+5% on top (RTX 3090: ~28.3 MH/s measured)
• Byte-exact same walk v2 — nothing changes for pools or nodes, your number goes up because the walk is faster
• Tip: the walk is memory-bound — memory OC gains a few extra %, core clock barely matters

Rebuilt for max GPU compatibility
• Fatbins regenerated with the new kernel: native sm_61→90 (Pascal → RTX 4090/Hopper) + PTX fallback — RTX 50 (Blackwell) JIT-compiles the optimized kernel too
• llama inference engine now multi-arch (sm_70→90 + PTX), identical in Linux zip & HiveOS tar
• Requirements unchanged: r535+ drivers, glibc ≥ 2.34

New: `--force-model` — per-GPU tier override (CUDA order, CSV): `--force-model light,very-high`. Only bypasses the local VRAM check — the tier is still proven by the walk on the real weights, so forcing is self-penalizing. Credit @ocminer!

Fixes
• Complete GGUFs on disk reused even if the `.ok` sentinel was lost — no more multi-GB re-downloads; partials no longer truncated (thanks @fearke85, #23)
• CUDA lib installer output now visible in the miner log (thanks @DragonKeeper, #22)
• no-AVX/BMI2 is now the default build — no more SIGILL on pre-Haswell CPUs, zero cost (thanks @CLusmi)

[lightbit]

Keryx (KRX) is now listed on Chainquiry

Keryx is an open-source Layer-1 BlockDAG designed to combine GPU mining with decentralised AI inference.

Project highlights:

• Ticker: KRX
• Category: AI / Layer 1 / BlockDAG
• Consensus: Proof-of-Model Proof of Work
• BlockDAG protocol: GHOSTDAG
• Network rate: 10 blocks per second
• Mining hardware: NVIDIA GPUs
• AI system: Optimistic Proof of Inference
• Status: Live

The project provides open-source node and mining software, a browser wallet, block explorer, BlockDAG visualiser and AI inference interface.

Explore the complete Keryx profile:
https://chainquiry.com/projects/keryx/

Official website:
https://keryx-labs.com/


Chainquiry is an independent cryptocurrency project directory. A listing is not an endorsement or investment recommendation. Always verify project information through official sources and conduct your own research.

[Keryx Labs]

Keryx-node-v1.4.2-OPoI

https://github.com/Keryx-Labs/keryx-node/releases/tag/v1.4.2-OPoI

Mandatory, node-only update. Nodes below v1.4.2 are rejected.

What happened
On July 30 nodes started disagreeing on coinbase rewards (COINBASE MISMATCH) and disqualifying each other's blocks. The root cause was a coin-age accounting bug that made long-running nodes slowly drift from freshly started ones, v1.4.2 fixes the root cause.

What's Changed
- PoM proofs now always travel with blocks served during sync (IBD).
- Self-healing proofs: a node missing a recent block's proof re-fetches it from its peers automatically.
- Proofs received during sync are verified before being stored.
- Fresh difficulty reset + new chain anchor (H5.4).
- IBD: inconclusive sync negotiations no longer get you banned.
- PoM proof store shrinks from ~6 GB to ~1.2 GB.
- Underpaid AI requests are rejected before they ever reach a block.
- Memory fixes (including the cause of OOMs on 16 GB machines).

How to update:
- Stop your node and replace the binary with v1.4.2.
- Delete your datadir.
- Restore the snapshot below (mandatory).
- Restart.

Snapshot
- https://huggingface.co/datasets/Keryx-Labs/datadir/resolve/main/datadir.zip
- https://drive.google.com/file/d/1yCi059fwtLalKxzC2P-f_3dJQPCnCzrZ/view?usp=sharing
- https://keryx-labs.com/datadir.zip.torrent
- https://keryx-labs.com/datadir.zip

Credits
@fearke85 — PR #23: storage profiling and memory-sizing work; found several caches sized in entries instead of bytes, including the one behind the OOM reports on 16 GB machines.
@ocminer — PR #25: the AiRequest validation rules that keep underpaid requests out of blocks; this release extends them to mempool admission.
@ryancole03 — PR #26: P2P robustness fix, merged as-is.
@nectopower — review of PR #25 that caught the rules being enforced in only one place, and the escrow bug reports behind miner v0.4.2.
Everyone who sent screenshots and logs on July 30 — the incident was caught thanks to you.

[Keryx Labs]

Keryx Miner v0.4.3 is out!

https://github.com/Keryx-Labs/keryx-miner/releases/tag/v0.4.3-OPoI

Escrow — claim 87 rewards in one transaction, one fee

- Escrow claims are no longer trickled out one by one: the miner now claims 87 escrow outputs in a single transaction.Claiming those same 87 rewards one at a time used to cost 87×0.3=26.1 KRX in fees, now it's 0.3 KRX!
- Full batches only: a claim goes out complete or not at all; partial-batch bisection remains as a short-lived repair path.
- The escrow state is validated against the node at boot: entries pointing at reorged or unknown blocks are purged instead of poisoning claims forever.
- If your miner went through the July 31 network wedge and has stuck claims: run once with --recover-escrow after upgrading. Two out of two affected miners we inspected had a polluted state; the one-shot recover fixes it.

Inference — repetition penalty

- The in-process llama engine now applies a repetition penalty (256-token window, 1.10) in the sampling chain. This fixes the looping generations some models (notably GLM-4-9B) produced on the livefeed.
- libkeryx-llama.so changed — always replace it together with the binary. Both ship in this package.

Telemetry & TUI (thanks @DragonKeeper, #24)

- NVML is now the primary telemetry source: GPU memory temperatures (GDDR6X), core temp, fan, power — with nvidia-smi only as fallback, and a HiveOS nvtool fallback for memory temps.
- Per-device Blocks Accepted / Rejected counters in the TUI and stats API; block logs now name the device (Found a block on GPU #1).
- Efficiency now displayed as hashrate per watt.

Stability (thanks @DragonKeeper, #25)

- Transient CUDA faults (illegal address & co.) at miner load now drop the stale state for the affected GPU only and rebuild on the next cycle, instead of hot-looping on a wedged runtime.

[Keryx Labs]

Keryx-node-v1.4.5-PoM

https://github.com/Keryx-Labs/keryx-node/releases/tag/v1.4.5-PoM

⚠️ Mandatory upgrade. H6 activates on mainnet at DAA 77,203,262 (≈ Sunday 16 August 2026, 18:00 CEST, ±~30 min). Nodes and miners not on v1.4.5 before the gate will fork off the network.

Proof-of-Model v3 — the matrix-state walk

The heart of H6. The possession proof moves from the hash-based walk to a matrix-state walk with a spot-check verifier, hardening Proof-of-Model against possession forgery.

- New prover + spot-check verifier with attack-case tests; wire layout and verification wired into check_pom_proof, gated at the H6 score.
- The proven tier is committed in the block header and read for reward and service-bond accounting.
- Difficulty-reset companion armed at the gate to absorb the new walk's hashrate profile; the difficulty window re-converges afterward.
- Witness handling — a bad possession witness invalidates the witness, never the block hash (no witness-poisoning).
- Service-bond — enforced inference SLA

A new system that makes miners actually serve inference, not just hold the model.

- Escrow-key identity + mandatory delegation cert — a miner's service identity is its payout key; the wallet/CLI (delegate-escrow) signs the cert that authorises it.
- Per-request cohort audit — every accepted AiRequest audits the full eligible tier cohort; token-scaled service windows with a max_tokens cap.
- Escalation — rate-limited strikes, 24 h production suspension (burns the miner cut, does not reject the block), and escrow burn for a finality-deep miss.
- Sealed service state — a MuHash commitment of the service ledger in headers at the pruning point, downloaded and verified at IBD, so a fresh node re-derives the exact same standings.
- Signed V2 AiResponses and era-gated payloads.

Rewards & lineup

- H6 model lineup — tier 0 Qwen3.5-9B, tier 1 GLM-4-9B, tier 2 Gemma-4-12B (16 GB cards), tier 3 Qwen3.6-27B, tier 4 Kimi-48B.
- Tier-reward schedule re-spaced to 10-point steps across the five tiers.
- Fully-burned miner/escrow outputs are omitted rather than emitted as zero-value; merged red-block subsidies burn at H6.

Fixes & hardening

- UTXO commitment — fixes a state re-derivation divergence during reorgs that could wedge a re-syncing node; upgraded nodes recover normally.
- Recovery flag — KERYX_REVALIDATE_DISQUALIFIED=1 heals a wedged node in place (re-validates disqualified chain blocks) instead of a full resync.
- Block relay — recent orphan roots now require their possession proof at the H6 gate; retry applies the same proof policy as the first attempt.
- IBD — unverified carried PoM tiers are discarded.
- Escrow claims — burned-escrow rejections name the affected outpoints so miners drop them directly.

New RPCs

getServiceStrikes and a lifetime strike tally per identity · GetUtxoEntriesByOutpoints (153) · GetUtxoCountByAddress (152). Plus chunked utxoindex scans and outbound peer target raised 8 → 16.

Acknowledgements

@ocminer — suggested the idea behind the Proof-of-Model v3 spot-check design and helped review its soundness.
@jaja (neuropool.net) — review of the block-relay path; surfaced the orphan-root proof-policy gap fixed in this release.
@ryancole03 — IBD PoM-tier authentication (PR #27) and the coin-age maturation reorg fix.
@izzback — spotted the witness-poisoning issue (invalidate the witness, never the block).
@gryph — spotted the merged red-block subsidy burn.

Thanks to every operator who reported issues and ran the testnet.

[Keryx Labs]

🚨 Mandatory Upgrade: Keryx Node v1.5.6
https://github.com/Keryx-Labs/keryx-node/releases/tag/v1.5.6-PoM

This release fixes the 2026-08-28 chain split and underlying sync bugs. All node operators (including pools) must update within 24 hours to avoid forking off.

Inference requests remain paused until the network converges.

🛠️ How to Upgrade (Pick one):
Option A — Snapshot (~5 min, recommended)

- Stop keryxd
- Delete your datadir (~/.keryx-labs/keryx-mainnet/datadir by default)
- Download and extract datadir.zip (17.7 GB)
👉 https://huggingface.co/datasets/Keryx-Labs/datadir/resolve/main/datadir.zip
👉 https://drive.google.com/file/d/1GtMzVPnxnLii1bPx5gRaOPYD6nFYOnDS/view?usp=sharing
👉 https://keryx-labs.com/datadir.zip
- Start keryxd v1.5.6

Option B — Resync from scratch (~1h15)

- Stop keryxd
- Run with --reset-db (or delete the datadir)
- Start directly on v1.5.6

📌 Key Highlights:
- Consensus & Rewards: Fixed reward mint window validation to eliminate node disagreement and chain splits
- Sealed Service State: State import is now mandatory and pinned to a mainnet checkpoint to prevent syncing on empty or partial states
- IBD & Peering: Proper recovery handling after long downtimes and a 24-hour ban for obsolete nodes (< 1.5.3)
- Protocol v11: Compact PoM proofs via Merkle multiproofs (~7.4% bandwidth saved per block, shoutout to @Fear )

[Keryx Labs]

What an intense afternoon on Keryx!

 1:00 PM: Discovery of an exploit by a team member

 3:00 PM: Exploit fixed, emergency hardfork scheduled 3 hours later

 5:00 PM: Release of the new node and new miner

 6:00 PM: H10 Hardfork successful, network is clean.

A massive thank you to the community for being so responsive in updating your setups in record time!

https://x.com/Keryx_Labs/status/2094171024728437218