The v0.2.12 on-chain upgrade pull request opened for community review on April 15, 2026. The release collects the largest set of protocol changes since Gonka MainNet launched in August 2025, including multi-model Proof of Compute, consensus-level transaction fees, a standalone devshard runtime, and every CertiK audit fix identified during the protocol review. This is a call for review, not a governance vote — formal voting begins once the review window closes.

Background

The centerpiece of v0.2.12 is multi-model Proof of Compute (#1039). Gonka currently runs a single governance-approved model across the entire network — Qwen3-235B-A22B. The new design splits PoC into per-model groups. Each approved model generates its own local PoC weight, and total consensus weight is aggregated across models via governance-set coefficients. The first additional model activates three epochs after the upgrade passes.

Consensus-level transaction fees land through #937 and #981. A governance-controlled gas price applies to generic transactions, while protocol-duty messages — PoC, validations, inference, BLS DKG — are exempt via a new NetworkDutyFeeBypassDecorator. The MsgPoCV2StoreCommit message carries a two-component fee (base validation plus a count-linear term) as the primary Sybil defence on the commit path. Host onboarding documentation is updated at docs/host_onboarding.md.

The devshard standalone runtime (#1045) decouples devshard releases from the DAPI and mainnet release cycles, letting shard operators ship updates independently. The audit fix bundle closes ten CertiK findings across PRs #1020, #1021, #1022, #987, #949, #988, #825, #1011, #1029, and #789. Additional protocol hardening covers a stronger PoC v2 RNG — moving from 32-bit entropy to full 256-bit entropy, activated through a separate governance vote — on-chain propagation of the mlnode version on each HardwareNode, DKG dealer consensus fixes, aligned legacy validator slashing under required-collateral semantics, atomicity of the devshard escrow fund, and zero-timestamp tolerance in inference_finished event parsing.

Why it matters for Gonka

Multi-model PoC is the change that opens the network to model diversity. Before v0.2.12, every host ran the same model and competed for one weight pool. Per-model groups let operators host different workloads against separate pools, and governance can approve new models — lightweight reasoning, domain-specific, or larger frontier models — without a full consensus migration each time. For regular users this means the Gonka SDK will eventually expose more than one model behind the same OpenAI-compatible interface.

Consensus-level fees change the economics in a narrower but important way. Gonka charged only for AI inference before v0.2.12. Generic transactions were free and therefore open to spam. The new fee model charges for blockspace while keeping all protocol-duty traffic exempt, preserving the zero-fee guarantee for inference itself. The two-component fee on MsgPoCV2StoreCommit targets the specific vector most useful to a Sybil attacker: cheap bulk PoC commits.

What's next

Review is open now. Community members who surface bugs, edge cases, or security issues may be eligible for bounties during the next upgrade cycle. Formal governance voting starts when the review window closes, and the binary rolls out through the standard on-chain upgrade proposal documented in /docs/upgrades.md.

Host operators preparing for the upgrade should plan four concrete actions once the proposal passes:

  1. Ensure the cold account holds enough balance — around 100 GNK — to cover the auto-granted fee allowance spend limit.
  2. For each governance-approved model, either deploy it, delegate it, or explicitly refuse it. The included additional model activates three epochs after the upgrade.
  3. Deploy the versiond service from docker-compose.yml using the latest commit on main.
  4. Recreate the proxy container with the new version and parameters documented in the upgrade guide.

No action is required from regular network users before the proposal lands. Wallets, API keys, and the OpenAI-compatible SDK clients keep working without changes. The transition happens at the host and validator layer.