Bitcoin Forum
May 27, 2026, 09:10:33 PM *
News: Latest Bitcoin Core release: 31.0 [Torrent]
 
   Home   Help Search Login Register More  
Pages: « 1 2 [3]  All
  Print  
Author Topic: [ANN] KERYX | Live AI Inference | 100% Fee Burn | GPU PoW BlockDAG | 10 BPS  (Read 963 times)
Keryx Labs (OP)
Newbie
*
Offline

Activity: 23
Merit: 0


View Profile
May 21, 2026, 07:45:18 PM
Last edit: May 21, 2026, 08:05:23 PM by Keryx Labs
 #41

Here is the latest keryx-miner version v0.2.7. Available for Windows, Linux and HiveOS.

https://github.com/Keryx-Labs/keryx-miner/releases/tag/v0.2.7-OPoI

v0.2.7 — Multi-Model Inference & IPFS Distribution

What's new:
Multi-model support The miner now supports two inference models selectable at launch:
DeepSeek-R1-8B GGUF — default, higher quality responses
TinyLlama-1.1B — lightweight mode for lower-end GPUs, enabled with --light
IPFS model distribution
Models are now downloaded via the IPFS gateway instead of HuggingFace. No account, no rate limits, fully decentralized from day one.

Fixes:
Escrow recovery — streamed the HTTP response body to bypass the 10 MB ureq limit that was silently failing on large recovery payloads
Chat templates — DeepSeek now uses the LLaMA-3 chat template, TinyLlama uses Zephyr — fixes malformed inference outputs

Internal:
Escrow claiming logs demoted to debug to reduce noise in production output


EDIT: If your GPU isn't powerful enough to run DeepSeek R1, simply add the --light flag to run the miner with TinyLlama only
Keryx Labs (OP)
Newbie
*
Offline

Activity: 23
Merit: 0


View Profile
May 24, 2026, 06:02:10 PM
Last edit: May 24, 2026, 06:23:50 PM by Keryx Labs
 #42

Keryx Hardfork — DAA 15,550,000

Estimated activation: Thursday, May 28, 2026 — ~14:00 UTC

---
Why This Hardfork

While monitoring the live network following the integration of DeepSeek-R1-8B alongside TinyLlama, we identified a consensus-level vulnerability.

Miners who haven't updated their software are still attempting to answer DeepSeek queries using incorrect or mismatched models. The results they submit are not verifiable as coming from the declared model. We cannot rely on miners voluntarily updating — economic incentives alone are insufficient when bad data can still earn rewards.

The only clean, permanent fix is to enforce model capabilities strictly at the consensus level, with model identity cryptographically tied to the actual weight files. This requires a mandatory hardfork.

---
Why We're Launching Now

We are aware this delays exchange listings by a few days. We made that call deliberately.

Shipping a hardfork after listing creates far more friction — coordinating node operators, miners, and exchanges simultaneously under time pressure. Doing it now, while the network is growing but not yet listed, gives everyone a clean runway. A few days of delay is a reasonable price for a solid, trustworthy consensus from day one.

Network integrity comes first. Always.


---
Inference Rewards — Real KRX Utility
 
Every AiRequest transaction carries a fee in $KRX set by the requester. When a miner answers the request and the result passes the challenge window unchallenged, that fee is released to the miner.

Protocol inference prices:

- TinyLlama-1.1B — 0.5 KRX + 0.3 KRX network fee = 0.8 KRX minimum,
- DeepSeek-R1-8B — 2.0 KRX + 0.3 KRX network fee = 2.3 KRX minimum,

Requesters can set a higher inference fee to prioritize their request — miners will naturally serve the highest-paying requests first.

Miners earn $KRX in two ways:
- Block reward — standard PoW coinbase, same as before,
- Inference fee — paid per completed AiRequest, released after the challenge window,

This is real utility for $KRX: it is the currency you pay to run AI on the network. Not speculation — demand for inference creates direct, measurable buy pressure. Every agent, every dApp, every user that calls an AI on Keryx spends $KRX.

With this hardfork, the AiResponse tx formally links the miner's identity, the committed result, and the reward claim in a single verifiable chain of transactions:

AiRequest tx  →  AiResponse tx  →  challenge window  →  fee released
   (fee locked)     (miner commits)    (1000 blocks)          (to miner)

---
What Changes
  
AiResponse as a Dedicated Transaction,

Miners no longer embed results in the coinbase. Instead, after completing an inference, they submit a standalone AiResponse transaction (subnetwork 0x04) with a fixed 78-byte payload:

[request_hash : 32 bytes]  — links back to the AiRequest tx
[challenge_window_end : 8] — DAA after which the result is final
[response_ipfs_cid : 34]   — IPFS CIDv0 of the result content
[response_length : 4]      — token count

  
Results Live on IPFS,

The actual text is uploaded to IPFS. Only the 34-byte CID is committed on-chain:

- Unlimited response length — no on-chain size constraint,
- Content-addressed — the CID cryptographically proves the content,
- Decentralized — any IPFS node can serve the result,

  
Model Identity Tied to Weight Files,

model_id is now sha2-256(primary_weight_file), extracted directly from the IPFS CID of the model weights. Anyone can verify:

download the CID from IPFS → sha256sum the file → compare to model_id

A miner cannot claim to run TinyLlama or DeepSeek-R1-8B with an arbitrary model_id. The identifier is now a cryptographic fact about the actual files.

OPoI Commitment Updated,

The on-chain commitment for fraud proofs is now response_ipfs_cid[2..34] — the SHA2-256 digest of the IPFS content. Challengers can point to exactly what content hash was committed.

Auto-managed IPFS Daemon,

Miners don't need to manually install IPFS. The miner binary automatically downloads kubo and starts it on first run. One new flag: --ipfs-url (default http://127.0.0.1:5001/).


---
What Miners Need to Do

1. Update to the new miner and node binaries before DAA block 15,550,000
2. Nothing else — kubo installs and starts automatically on first run,

---
What Doesn't Change

- Block structure, PoW, GHOSTDAG — untouched,
- Escrow / OPoI Phase 1–2 — fully compatible,
- All existing AiRequest transactions — still valid,
- Pre-hardfork blocks — still indexed correctly,

The updated binaries for both the node and the miner will be available on GitHub starting tomorrow.
Keryx Labs (OP)
Newbie
*
Offline

Activity: 23
Merit: 0


View Profile
May 26, 2026, 04:36:10 PM
 #43

Keryx Node v1.2.1

https://github.com/Keryx-Labs/keryx-node/releases/tag/v1.2.1-OPoI

Hardfork release — This version contains consensus changes incompatible with v1.2.0. Update is mandatory to stay on the network.

  ---
  Highlights

  - UTXO Escrow for inference rewards — inference_reward is now flows through a consensus-validated UTXO escrow, ensuring only legitimate inferences get paid out.
  - Model Cap Enforcement — The consensus now rejects any AiResponse whose model_id is not declared in the coinbase ai:cap: field. Miners must advertise their supported models.
  - Fixed 78-byte AiResponse layout — AiResponse is now a fixed 78-byte structure with an encoded IPFS CIDv1. This breaks binary compatibility with v1.2.0.
  - New supported models — DeepSeek-R1-32B and LLaMA-3.3-70B added, with tiered rewards: 0.5 / 1.5 / 2.5 / 4.0 KRX based on model size.

  ---
  Consensus & OPoI

  - Token-proportional minimum reward: 0.05 KRX per 64 tokens generated
  - model_id = sha2-256(weight_file) derived from the IPFS CIDv1 — deterministic and verifiable off-chain
  - Per-model reward minimums + priority_fee payout to miners
  - Hardfork activation is now logged at the DAA score trigger

  Mempool

  - AiResponse and AiChallenge transactions are exempt from the minimum fee check — they carry their own inference_reward

  P2P / Network

  - Auto-ban phantom nodes after repeated ping timeouts
  - Ban peers on IBD spam and finality violations

  Fixes

  - UPnP: port mapping failure downgraded from WARN to INFO with an actionable message — UPnP is absent on most VPS routers and is not a critical error

  ---

The node will automatically detect the hardfork at the scheduled DAA score (15,550,000, estimated activation on Thursday, May 28, 2026, at ~14:00 UTC) and switch to the new consensus rule
Keryx Labs (OP)
Newbie
*
Offline

Activity: 23
Merit: 0


View Profile
Today at 07:01:27 PM
 #44

keryx-miner v0.2.8 — GPU Inference OPoI (Full Pipeline)

https://github.com/Keryx-Labs/keryx-miner/releases/tag/v0.2.8-OPoI

This release delivers the complete Optimistic Proof of Inference pipeline: models now run natively on CUDA, results are published to IPFS, and the miner submits an on-chain AiResponse transaction — all automatically, on every mined block.

What's new

Real GPU inference (CUDA 12.6)
Models are now loaded directly onto the GPU. The miner verifies VRAM budget at startup and places models accordingly — no more silent CPU fallback.

Full OPoI pipeline end-to-end
Mining template arrives → GPU generates a response → result is uploaded to IPFS → an AiResponse transaction (subnetwork 0x04) is broadcast on-chain with the IPFS CID → inference_reward is auto-claimed via escrow UTXO.

Multi-model tiers

--light — TinyLlama 1.1B only, any GPU with 6 GB+ VRAM
(default) — TinyLlama + DeepSeek-R1-8B, RTX 3060 12 GB / 3070 / 3080
--high — adds DeepSeek-R1-32B, RTX 3090 / 4090 (24 GB+)
--very-high — adds LLaMA-3.3-70B, RTX 5090 (32 GB+)
New models: DeepSeek-R1-32B and LLaMA-3.3-70B
Both distributed via IPFS, downloaded on demand at startup.

Model prefetch before mining
Model files are fetched from IPFS before the miner starts looking for blocks — no missed inference jobs on first launch.

Kubo auto-download
The IPFS node (Kubo) is downloaded automatically at runtime if absent — no manual setup required.

Capability announcement
The miner broadcasts its loaded models in the coinbase extra_data (/ai:cap:<model_id>/) so the network knows which models are available before routing requests.

Better response quality
Repeat-penalty sampler (1.15 / 64-token window), multi stop-token support, blocks stripped before publishing to IPFS, stronger Keryx identity system prompt.

Escrow fixes
Inference reward claims are prioritised over coinbase to prevent starvation. Seq-lock cooldown added for retry loops.
Pages: « 1 2 [3]  All
  Print  
 
Jump to:  

Powered by MySQL Powered by PHP Powered by SMF 1.1.19 | SMF © 2006-2009, Simple Machines Valid XHTML 1.0! Valid CSS!