What This Release Is About
Gonka v0.2.11 shipped on March 19, 2026 with 18 merged pull requests. The update covers five major areas: a new off-chain inference architecture, performance improvements to block execution, security patches from the CertiK audit, a reworked collateral and slashing system, and cross-chain trading via IBC. There is also a bounty payout of 157,750 GNK to contributors.
The release has been deployed and verified on testnet. An on-chain governance proposal will activate it for all validators on MainNet.
Inference Shards: Moving Coordination Off-Chain
The most significant architectural change in this release is subnet-based inference shards (PR #877).
Until now, every inference request was fully coordinated on-chain — the chain tracked who runs what, managed scheduling, and handled settlement for each individual request. That works, but it puts a ceiling on throughput because every inference adds transactions to the block.
The new approach splits this into two layers. A lightweight subnet handles the real-time coordination: assigning work to nodes, managing transport between them, handling cryptographic signing, and storing intermediate state. The chain only steps in at the beginning (session setup and escrow) and the end (settlement). Everything in between happens off-chain.
The subnet subsystem includes five new components: state management, peer-to-peer transport, threshold signing, local storage, and API integration. Each subnet operates as a self-contained coordination unit for a group of inference nodes.
This is currently marked experimental with limited access. The full architecture proposal is documented in the proposals/inference/ directory on GitHub. The long-term goal is to remove the per-inference bottleneck on the chain and allow the network to scale horizontally — more subnets means more parallel inference capacity without increasing block size.
Performance: Four PRs Targeting Hot Paths
Four separate pull requests focused on reducing overhead in the most frequently executed code paths:
StartInference / FinishInference (PR #812) — These two messages are called for every inference request. The PR removes unnecessary writes to chain state during the inference lifecycle and reduces the number of queries executed. Previously, each inference would write intermediate stats that could be computed lazily. Now those writes only happen when actually needed.
Key verification (PR #779) — The inference message path included cryptographic signature verification at multiple stages. Some of these checks were redundant because the protocol already guarantees validity at earlier stages. This PR identifies where those guarantees hold and skips the redundant verification, reducing CPU work per inference message.
MsgValidation and MsgClaimRewards (PR #874) — Validators call these messages frequently. The PR adds transient caching so repeated lookups within the same block hit memory instead of storage. It also restructures the validation logic to avoid redundant computation and introduces state pruning support to keep storage from growing indefinitely.
Unified permissions (PR #760) — Authorization checks in the inference module were scattered across multiple files with duplicated logic. This PR consolidates them into a single permissions layer, making the behavior explicit and easier to test. It also avoids repeated permission lookups within the same transaction.
The combined effect is lower per-block execution time, which directly improves block production stability under heavy inference load.
Security: CertiK BLS Audit and CometBFT Patch
The CertiK audit of the BLS (Boneh-Lynn-Shacham) cryptographic module identified issues in two areas, both fixed in PR #822:
Threshold validation — The distributed key generation (DKG) protocol had a flaw where the threshold for accepting key shares was not correctly enforced under certain edge cases. An attacker with control of enough validators could potentially manipulate the group key. The fix tightens the threshold check to match the protocol specification.
Duplicate slot handling — During threshold signing, if a validator submitted signatures for the same slot multiple times, the aggregation logic could count them twice. This has been corrected to deduplicate before aggregation.
Separately, CometBFT (the consensus engine) was updated to v0.38.21 (PR #675) to patch the CSA-2026-001 vulnerability — a consensus-level advisory disclosed in the upstream CometBFT project.
The application.db bloat issue (PR #867) was also resolved. Many node operators had reported their database growing much faster than expected. The root cause was insufficient pruning of historical application state. The fix introduces proper pruning, and operators who were affected should see their database size stabilize.
Collateral and Slashing: Fairer Economics
Three PRs reworked how collateral deposits interact with slashing penalties and mining rewards:
Proportional slashing (PR #868) — Previously, if a validator was slashed, the penalty was calculated against their total deposited collateral. This meant a validator who deposited 10x the minimum requirement would lose 10x more than one who deposited exactly the minimum, even for the same offense. The new logic calculates the penalty based only on the required collateral amount. Extra deposits are protected from slashing, which removes a disincentive to over-collateralize.
Collateral-aware rewards (PR #888) — Miners who deposited less than the required collateral were still receiving full mining rewards. The fix scales their effective earning power in proportion to their actual collateral relative to the requirement. If you deposit 50% of the required amount, your effective weight (and therefore rewards) drops accordingly.
Slashed coins to governance (PR #775) — Previously, slashed collateral was burned (removed from circulation). Now it is redirected to governance-controlled destinations. This means the network retains the value and can redistribute it through community proposals instead of permanently destroying it.
IBC Trading: Cross-Chain Assets
PR #814 introduces the ability to trade assets from other Cosmos chains on the Gonka network. This is built on top of the Inter-Blockchain Communication (IBC) protocol that was already supported for transfers.
The implementation adds new chain-level messages and queries for trading IBC-denominated tokens. The community sale and liquidity pool CosmWasm contracts have been updated to support these new asset types. Trading pairs must be approved through governance — the system does not allow arbitrary token listing.
This makes Gonka a more complete DeFi environment: users can now bridge assets from other Cosmos chains and trade them natively without leaving the Gonka ecosystem.
Other Changes
-
Batch Transfer with Vesting (PR #835) — Previously, sending vesting tokens to multiple addresses required submitting a separate governance proposal for each recipient. This was impractical for bounty distributions. The new batch transfer lets a single proposal send to many addresses at once.
-
Delete Governance Model (PR #773) — Adds a transaction type for removing a governance model from the system. The deletion is designed to not affect the current epoch, so active inference continues uninterrupted.
-
Data Race Fixes (PR #543) — A comprehensive review identified and fixed all concurrent data access issues across the codebase. This reduces the risk of non-deterministic behavior under load.
-
Reproducible Sampling Protection (PR #719) — Stage 1 of the sequence check system that ensures inference outputs are reproducible across different nodes, preventing validators from returning arbitrary results.
-
Updated CONTRIBUTING.md and issue templates for clearer community contribution guidelines.
Bounty Distribution: 157,750 GNK
The release includes 24 bounties paid to contributors:
| Contributor | Amount | Work |
|---|---|---|
| @qdanik | 27,000 GNK | vLLM 0.15.1 compatibility — basis for next ML node |
| Axel-t, @Red-Caesar | 25,000 GNK | PoC integration into vLLM v0.11.1 |
| @clanster, @baychak | 22,500 GNK | vLLM 0.15.1 compatibility experiments |
| @akup + @x0152 | 15,000 GNK | Release management |
| @Lelouch33 | 7,500 GNK | application.db fix with full testing |
| @huxuxuya | 5,000 GNK | Batch Transfer with Vesting |
| @qdanik | 5,000 GNK | Collateral slashing vulnerability |
| @qdanik | 5,000 GNK | Wind-down window vulnerability |
| @ouicate | 10,750 GNK | Three security reports (DoS, state bloat, bridge parsing) |
| @blizko | 11,000 GNK | HTTP 502 griefing + dust transaction vulnerability |
| Others | 23,000 GNK | Node snapshot solving, StartInference fixes, reviews |
The vLLM 0.15.1 compatibility work (combined 49,500 GNK) is the largest bounty category — this work lays the foundation for the next generation ML node that will support newer model architectures.
Migration
The on-chain migration performs four operations: disables ClaimValidationEnabled, rebuilds active participant caches for the current and previous epoch, converts epoch-group validations to the new entry-based format, and runs the community-sale CosmWasm contract migration.
Existing hosts do not need to manually update their containers. Binary versions will be updated through the on-chain upgrade proposal. The new container versions in the repository are for hosts joining after the upgrade.
Source: GitHub release/v0.2.11