[ANN] KERYX [KRX] | GPU PoW // 10 BPS // Verifiable AI Inference on BlockDAG

[Keryx Labs]

Keryx is not just another PoW coin. It is a decentralized infrastructure layer for verifiable AI.

We are building a BlockDAG-based network where mining serves a dual purpose: securing the ledger and providing verifiable compute power for Small Language Models (SLM).

🧠 The Core Thesis
Current AI inference is either closed-source and centralized or lacks cryptographic proof of execution. Keryx changes this by integrating Proof-of-Inference (OPoI) directly into the consensus mechanism. Miners do not just solve hashes; they execute deterministic AI tasks that the network can validate.

🛠 Technical Specifications

Consensus: BlockDAG (High throughput, low latency).

Mining Algorithm (KeryxHash): A specialized PoW evolved from kHeavyHash, optimized for GPU efficiency:

ASIC Isolation: Implements KERYX_MATRIX_SALT (32-byte domain-specific salt) to isolate the network from existing ASICs.

WaveMix Technology: A 4-round ARX post-processing step increasing GPU register pressure to favor consumer hardware.

Security Mutation: Re-engineered consensus layers to ensure unique, cryptographically secure block signing.

On-Chain Inference: Native integration of TinyLlama-1.1B. Every query is processed by miners and validated on-chain via OPoI.

Core Language: Rust (Performance & Memory safety).

Distribution: Fair Launch. No premine, no ICO, no VC allocations. To ensure long-term development, the protocol uses a 5% R&D allocation on every block reward, dedicated to network infrastructure and core engine improvements.

🛠 The Keryx Ecosystem

Explorer: A custom-built BlockDAG visualizer. Trace transactions, monitor OPoI tags, and verify SLM execution in real-time.

Inference API: High-performance API built in Rust. Provides a robust interface for third-party developers to query the Keryx network directly.

Web Wallet: A sleek, non-custodial Next.js web interface. Designed for secure in-browser key management.

🚀 Why Keryx?

Verifiable Compute: The result of an AI inference is tied to the block reward. If the inference isn't accurate, the block isn't valid.

Build in Public: We don't talk about moon. We talk about slm.rs, candle-core benchmarks, and throughput.

Community: This project is built for miners, engineers, and researchers. No hype, just code.

💻 Resources

GitHub: https://github.com/Keryx-Labs/

Website: https://keryx-labs.com/

Discord (Miners & Devs only): https://discord.gg/U9eDmBUKTF

X: https://x.com/Keryx_Labs

We are looking for early node operators and GPU miners. If you know how to compile from source and run a node on Linux, join the technical channel.

[Keryx Labs]

We've just launched the testnet. I'm not sure yet when we'll launch the mainnet; maybe this weekend

[WestHamFan10]

The OPoI challenge mechanism is interesting — if a miner submits a bad inference result and nobody challenges it within the dispute window, does the reward confirm anyway, or is there a minimum challenge period enforced at consensus level? Also curious how you're handling the cold-start problem before real inference demand exists — are the synthetic tasks from block hashes just for testing or will they persist on mainnet as a fallback

[Keryx Labs]

Great questions! Here's the honest breakdown of where we stand:

Challenge window & Staking:
We're currently in Phase 1, where validation is deterministic — every node recomputes the MLP result itself, so no dispute window is needed yet. The full optimistic model (miner stake, slashing, Groth16 fraud-proofs) is fully designed and spec'd for Phase 2. Once activated via a scheduled hard fork, rewards will be held pending until the challenge window elapses, with ZK verification triggering only if a dispute is raised.

Cold-start / Synthetic tasks:
This is already live and mandatory from block 1. The miner uses the block nonce as MLP input, and the node rejects any block without a valid tag — so miners are already running inference on every single block regardless of organic demand. Furthermore, TinyLlama-1.1B is already integrated: you can submit a prompt on-chain via a standard TX, the miner picks it up, runs real inference, and embeds the response in the next coinbase. We've already validated this end-to-end. Think of synthetics as the permanent protocol floor, with real AI queries layering on top as paid, higher-priority tasks.

To sum up: Validation is hard-enforced and TinyLlama is live today. The economic layer (stake/slash) will ship as a scheduled network upgrade — following the same proven pattern Ethereum used to add slashing post-launch.
Keryx's mainnet will officially launch this Sunday at 13:00 UTC

Get ready to try out AI inference on BlockDAG with TinyLlama-1.1B

Let's make history !

[_ili_]

an ideal model to start ai inference network, at least the gpus will be useful

[Keryx Labs]

Absolutely. Using TinyLlama-1.1B for the launch is a strategic move: it's the perfect balance between accessibility for miners and real-world complexity to validate our inference layer. The goal is to prove that the BlockDAG can serve for much more than just a store of value.

[@alymwehrli]

Another really cool project, congratulations to the development team and the designers, the website looks great. It's a cool idea within the AI ​​category, it has potential, it just needs more information about the company structure and the partners. It doesn't need to be a huge structure, it can be lean, but it's good to have a business structure. Feel free to contact me if you need help setting up a business structure.

[fdz]

Any tutorials for dummies on mining on Windows?   

[Laheeboo]

Install ubuntu 24.04 from windows store and you can run the node and the miner you don't have to have ubuntu operating system to do so.

[COBRAS]

Any tutorials for dummies on mining on Windows?   

Install node, open wallet and get adress at official www,run miner.All errors in installation AI answers help. Halt of workers - use gpu >= 4xxx or use command line to miner --blsbla and resolve for old gpu.

[Delaforetnoire]

an ideal model to start ai inference network, at least the gpus will be useful

With a lighter model, the GPU won’t be fully utilized, but for a more standard or ideal model, the GPU will be properly put to use.

[COBRAS]

update ballance refrashes !!! coins at ballance.

I mined a blocks but ballance is 0 !!!

Address
Address
keryx:qzs9f7yyrrhxzw3709glg5aeuavrlyvfpfdrhzn3uraluqdnhdnhgzjx789xe
Balance
0 KRX
Indexed txs
6
Transaction History
1–6 of 6
DAA Score
Dir
TX ID
Amount (KRX)
#2,549,761
IN
308dc72a5372589be351…cd8904
+5.292
#2,549,720
IN
bf4942dcd083dda4f06c…477419
+5.292
#2,549,390
IN
471148950ee5bf0c6389…a99066
+5.292
#2,547,704
IN
3d2a9b1a3d2129b972fe…175054
+5.292
#2,546,631
IN
ee5b04bbd9936490f756…347157
+5.292
#2,546,498
IN
0dd1fc3f218f1e666a10…c17a58

[Keryx Labs]

Checking your wallet, I can see that everything looks fine:


Address
keryx:qzs9f7yyrrhxzw3709glg5aeuavrlyvfpfdrhzn3uraluqdnhdnhgzjx789xe
Balance
2078.4549 KRX
Indexed txs
396

update ballance refrashes !!! coins at ballance.

I mined a blocks but ballance is 0 !!!

Address
Address
keryx:qzs9f7yyrrhxzw3709glg5aeuavrlyvfpfdrhzn3uraluqdnhdnhgzjx789xe
Balance
0 KRX
Indexed txs
6
Transaction History
1–6 of 6
DAA Score
Dir
TX ID
Amount (KRX)
#2,549,761
IN
308dc72a5372589be351…cd8904
+5.292
#2,549,720
IN
bf4942dcd083dda4f06c…477419
+5.292
#2,549,390
IN
471148950ee5bf0c6389…a99066
+5.292
#2,547,704
IN
3d2a9b1a3d2129b972fe…175054
+5.292
#2,546,631
IN
ee5b04bbd9936490f756…347157
+5.292
#2,546,498
IN
0dd1fc3f218f1e666a10…c17a58

I found the cause: the balance only updates after 1,000 confirmations. I definitely need to change that lol

update ballance refrashes !!! coins at ballance.

I mined a blocks but ballance is 0 !!!

Address
Address
keryx:qzs9f7yyrrhxzw3709glg5aeuavrlyvfpfdrhzn3uraluqdnhdnhgzjx789xe
Balance
0 KRX
Indexed txs
6
Transaction History
1–6 of 6
DAA Score
Dir
TX ID
Amount (KRX)
#2,549,761
IN
308dc72a5372589be351…cd8904
+5.292
#2,549,720
IN
bf4942dcd083dda4f06c…477419
+5.292
#2,549,390
IN
471148950ee5bf0c6389…a99066
+5.292
#2,547,704
IN
3d2a9b1a3d2129b972fe…175054
+5.292
#2,546,631
IN
ee5b04bbd9936490f756…347157
+5.292
#2,546,498
IN
0dd1fc3f218f1e666a10…c17a58

[vegeta12]

About the OPoI (Proof-of-Inference) implementation. Given the 10 BPS BlockDAG (which is a Kaspa fork, if I'm not mistaken), how are you solving the state continuity problem? Standard BlockDAG only validates transaction dependencies. With TinyLlama-1.1B natively integrated, the state of the AI model itself becomes part of the ledger. Does every block contain a full state snapshot of the model after executing the queries, or are you using a different mechanism for optimistic proofs? This is a massive storage and bandwidth concern at 10 BPS.

[Keryx Labs]

Good question — and yes, it's a Kaspa fork (BlockDAG, 10 BPS, KeryxHash PoW).

What's live on mainnet today:

Every block requires an OPoI tag in the coinbase extra_data. The miner loads TinyLlama-1.1B-Chat (via candle-transformers), takes the block nonce as input seed, generates a response, and commits an 8-byte hash of it. Blocks
without a valid tag are rejected by the node.

Users can submit inference requests on-chain via a TX payload (KRX:AI:1:{model, prompt, max_tokens}). The miner detects pending requests in the block template, runs TinyLlama on the prompt, and embeds the result
(base64) in the next coinbase extra_data. The indexer links request → response on-chain.

On the state continuity question:

There is no model state between blocks. TinyLlama's weights are fixed — each inference is a pure function of (weights + input). No checkpoint per block, no evolving state. The ledger records request/response hashes, not
activations or weights. At 10 BPS, this is ~200-400 bytes per inference in the coinbase payload.

What's next (Phase 2/3):
- Collateral + challenge window + ZK fraud proofs (Groth16)
- IPFS for model weight distribution (currently downloaded from HuggingFace)

[COBRAS]

Good question — and yes, it's a Kaspa fork (BlockDAG, 10 BPS, KeryxHash PoW).

What's live on mainnet today:

Every block requires an OPoI tag in the coinbase extra_data. The miner loads TinyLlama-1.1B-Chat (via candle-transformers), takes the block nonce as input seed, generates a response, and commits an 8-byte hash of it. Blocks
without a valid tag are rejected by the node.

Users can submit inference requests on-chain via a TX payload (KRX:AI:1:{model, prompt, max_tokens}). The miner detects pending requests in the block template, runs TinyLlama on the prompt, and embeds the result
(base64) in the next coinbase extra_data. The indexer links request → response on-chain.

On the state continuity question:

There is no model state between blocks. TinyLlama's weights are fixed — each inference is a pure function of (weights + input). No checkpoint per block, no evolving state. The ledger records request/response hashes, not
activations or weights. At 10 BPS, this is ~200-400 bytes per inference in the coinbase payload.

What's next (Phase 2/3):
- Collateral + challenge window + ZK fraud proofs (Groth16)
- IPFS for model weight distribution (currently downloaded from HuggingFace)

Why two days ago I mine a 10000/24h, now only 7000/24h?

Thank you.

I found the cause: the balance only updates after 1,000 confirmations. I definitely need to change that lol

update ballance refrashes !!! coins at ballance.

I mined a blocks but ballance is 0 !!!

Address
Address
keryx:qzs9f7yyrrhxzw3709glg5aeuavrlyvfpfdrhzn3uraluqdnhdnhgzjx789xe
Balance
0 KRX
Indexed txs
6
Transaction History
1–6 of 6
DAA Score
Dir
TX ID
Amount (KRX)
#2,549,761
IN
308dc72a5372589be351…cd8904
+5.292
#2,549,720
IN
bf4942dcd083dda4f06c…477419
+5.292
#2,549,390
IN
471148950ee5bf0c6389…a99066
+5.292
#2,547,704
IN
3d2a9b1a3d2129b972fe…175054
+5.292
#2,546,631
IN
ee5b04bbd9936490f756…347157
+5.292
#2,546,498
IN
0dd1fc3f218f1e666a10…c17a58

Ballance is come to my screen, so, I not wary more. Situation  normal now. Thanks. )

[chafer99]

Any tutorials for dummies on mining on Windows?   

Install node, open wallet and get adress at official www,run miner.All errors in installation AI answers help. Halt of workers - use gpu >= 4xxx or use command line to miner --blsbla and resolve for old gpu.

HI, i installed node and open wallet, running miner with command: keryx-miner --mining-address keryx:qqjjh8438x0zjg32f24nu... --port 22110
Miner displays an error:
"Thread <unnamed> panicked at plugins\cuda\src\lib.rs:68:59:
"called Results ::unwrap() on an Err value: NotSupported"
"fatal runtime error: Rust cannot catch foreign exceptions, aborting"

I have only AMD GPU, CPU.

Thank you

[Keryx Labs]

🔴 [NODE UPDATE] Auto-ban for invalid coinbase + main node incident

What happened last night
The main node diverged onto its own fork. Root cause: some participants modified their node software to remove the 2% R&D allocation from their coinbase. This caused enough invalid blocks to accumulate that the node's
virtual chain diverged from the network majority.

The node will be restarted and will be back in sync soon

---
New protection merged into master

We've just merged automatic banning of peers who relay blocks without the R&D allocation.

How it works:
Before a relayed block enters the consensus pipeline, the node checks that the coinbase contains the R&D output
If it doesn't → the relaying peer is immediately banned
Banned peers are not re-added to the known address list
The invalid block is dropped and never propagated further

The result: nodes and miners cheating the R&D allocation progressively get isolated from the network.

---
What you need to do

If you're running a keryxd node, please update:
git pull origin master
cargo build --release

 Restart keryxd with -reset-db to avoid a ban and --addpeer=144.91.120.224 to make sure you're on the right network:

./keryxd --utxoindex --rpclisten-json=default --rpclisten-borsh=default --reset-db --addpeer=144.91.120.224


⚠️  Nodes that don't update will continue relaying invalid blocks and will be banned by updated nodes for 24 hours

Thanks for your support 🙏

[COBRAS]

🔴 [NODE UPDATE] Auto-ban for invalid coinbase + main node incident

What happened last night
The main node diverged onto its own fork. Root cause: some participants modified their node software to remove the 2% R&D allocation from their coinbase. This caused enough invalid blocks to accumulate that the node's
virtual chain diverged from the network majority.

The node will be restarted and will be back in sync soon

---
New protection merged into master

We've just merged automatic banning of peers who relay blocks without the R&D allocation.

How it works:
Before a relayed block enters the consensus pipeline, the node checks that the coinbase contains the R&D output
If it doesn't → the relaying peer is immediately banned
Banned peers are not re-added to the known address list
The invalid block is dropped and never propagated further

The result: nodes and miners cheating the R&D allocation progressively get isolated from the network.

---
What you need to do

If you're running a keryxd node, please update:
git pull origin master
cargo build --release

 Restart keryxd with -reset-db to avoid a ban and --addpeer=144.91.120.224 to make sure you're on the right network:

./keryxd --utxoindex --rpclisten-json=default --rpclisten-borsh=default --reset-db --addpeer=144.91.120.224


⚠️  Nodes that don't update will continue relaying invalid blocks and will be banned by updated nodes for 24 hours

Thanks for your support 🙏

Rebuild from node sources is anoth for this ?
Again, mained a LOT blocks, but ballance not grow !

[2026-05-02T01:16:07Z INFO  keryx_miner::pow] Found a block: 80a26341db2f9bc008c6ae573df1a24c005863d06d469e42a4cd7cbe9c6d5d35 (Timestamp: 2026-05-02 01:16:07)
[2026-05-02T01:16:07Z INFO  keryx_miner::client::grpc] block submitted successfully!
[2026-05-02T01:16:13Z INFO  keryx_miner::miner] Current hashrate is 767.40 Mhash/s
[2026-05-02T01:16:13Z INFO  keryx_miner::miner] Device #0 (NVIDIA GeForce RTX 3090): 767.40 Mhash/s
[2026-05-02T01:16:23Z INFO  keryx_miner::miner] Current hashrate is 767.40 Mhash/s
[2026-05-02T01:16:23Z INFO  keryx_miner::miner] Device #0 (NVIDIA GeForce RTX 3090): 767.40 Mhash/s
[2026-05-02T01:16:23Z INFO  keryx_miner::pow] Found a block: 44a19dab48cfe3e2052279c3f7f6a3333b770bde09d565ad0c8d90f0cd42349c (Timestamp: 2026-05-02 01:16:23)
[2026-05-02T01:16:23Z INFO  keryx_miner::client::grpc] block submitted successfully!
[2026-05-02T01:16:33Z INFO  keryx_miner::miner] Current hashrate is 767.40 Mhash/s
[2026-05-02T01:16:33Z INFO  keryx_miner::miner] Device #0 (NVIDIA GeForce RTX 3090): 767.40 Mhash/s
[2026-05-02T01:16:43Z INFO  keryx_miner::miner] Current hashrate is 773.85 Mhash/s
[2026-05-02T01:16:43Z INFO  keryx_miner::miner] Device #0 (NVIDIA GeForce RTX 3090): 773.85 Mhash/s
[2026-05-02T01:16:53Z INFO  keryx_miner::miner] Current hashrate is 767.40 Mhash/s
[2026-05-02T01:16:53Z INFO  keryx_miner::miner] Device #0 (NVIDIA GeForce RTX 3090): 767.40 Mhash/s
[2026-05-02T01:17:01Z INFO  keryx_miner::pow] Found a block: 54a36dd4f59ac27806124071d0e5fed0d84201413ca17bd0adebf6a9747cab5e (Timestamp: 2026-05-02 01:17:01)
[2026-05-02T01:17:01Z INFO  keryx_miner::client::grpc] block submitted successfully!
[2026-05-02T01:17:03Z INFO  keryx_miner::miner] Current hashrate is 767.40 Mhash/s
[2026-05-02T01:17:03Z INFO  keryx_miner::miner] Device #0 (NVIDIA GeForce RTX 3090): 767.40 Mhash/s
[2026-05-02T01:17:05Z INFO  keryx_miner::pow] Found a block: 31a02dcea0ee46d3215e5169939425fc128fa7393d319d5491a64ecb47d2f1a4 (Timestamp: 2026-05-02 01:17:05)
[2026-05-02T01:17:05Z INFO  keryx_miner::client::grpc] block submitted successfully!
[2026-05-02T01:17:13Z INFO  keryx_miner::miner] Current hashrate is 773.85 Mhash/s
[2026-05-02T01:17:13Z INFO  keryx_miner::miner] Device #0 (NVIDIA GeForce RTX 3090): 773.85 Mhash/s
[2026-05-02T01:17:23Z INFO  keryx_miner::miner] Current hashrate is 767.40 Mhash/s
[2026-05-02T01:17:23Z INFO  keryx_miner::miner] Device #0 (NVIDIA GeForce RTX 3090): 767.40 Mhash/s
[2026-05-02T01:17:24Z INFO  keryx_miner::pow] Found a block: 512afe589d53471d99e96e294b4b25e2f26ff999a819537dcd6ff4d6c92aaf1c (Timestamp: 2026-05-02 01:17:24)
[2026-05-02T01:17:24Z INFO  keryx_miner::client::grpc] block submitted successfully!
[2026-05-02T01:17:24Z INFO  keryx_miner::pow] Found a block: b967f8edcf2dc0dbcb6e81f0e9254d86dce660e768a410a5ad33eca0217da30d (Timestamp: 2026-05-02 01:17:24)
[2026-05-02T01:17:24Z INFO  keryx_miner::client::grpc] block submitted successfully!
[2026-05-02T01:17:33Z INFO  keryx_miner::miner] Current hashrate is 767.40 Mhash/s
[2026-05-02T01:17:33Z INFO  keryx_miner::miner] Device #0 (NVIDIA GeForce RTX 3090): 767.40 Mhash/s
[2026-05-02T01:17:41Z INFO  keryx_miner::pow] Found a block: 83a885ca19ac2952c06909f46234938576a2e67fcff4c7f4e7005ba4d4720276 (Timestamp: 2026-05-02 01:17:41)
[2026-05-02T01:17:41Z INFO  keryx_miner::client::grpc] block submitted successfully!
[2026-05-02T01:17:43Z INFO  keryx_miner::miner] Current hashrate is 767.40 Mhash/s
[2026-05-02T01:17:43Z INFO  keryx_miner::miner] Device #0 (NVIDIA GeForce RTX 3090): 767.40 Mhash/s
[2026-05-02T01:17:47Z INFO  keryx_miner::pow] Found a block: fdbf68049a033bea7c82c9f0b2bf92cf8cd8836a1581e58c66efdc8de8215a1c (Timestamp: 2026-05-02 01:17:47)
[2026-05-02T01:17:47Z INFO  keryx_miner::client::grpc] block submitted successfully!
[2026-05-02T01:17:52Z INFO  keryx_miner::pow] Found a block: 950e94e3236934424af58bf37db1b46a694c1a657875d72c75bd88919d47f87f (Timestamp: 2026-05-02 01:17:52)
[2026-05-02T01:17:52Z INFO  keryx_miner::client::grpc] block submitted successfully!
[2026-05-02T01:17:53Z INFO  keryx_miner::miner] Current hashrate is 767.40 Mhash/s
[2026-05-02T01:17:53Z INFO  keryx_miner::miner] Device #0 (NVIDIA GeForce RTX 3090): 767.40 Mhash/s

-=========

71578.31241421   this frezed number of ballance more 3 hours with this cool  blocks findings!!! Resolthe this problem please ?

[Keryx Labs]

Yeah, need to rebuild the node. I know about the balance lag, I'll fix it ASAP!