The Gonka documentation now reflects the network's live multi-model setup. PR #1165, merged on 2026-06-04 by @patimen, rewrote the multi-model Proof of Compute guide (docs/host/multi_model_poc.md) so it matches what MainNet actually runs after the v0.2.13 upgrade, replacing an older snapshot that only described a single Kimi model.

Proof of Compute (PoC) is how Gonka reaches consensus: instead of solving hash puzzles, validators run real AI inference, and that work is what earns them weight on the chain. Multi-model PoC means the network no longer runs just one model. It runs several, and each model carries its own tuning parameters.

What changed

The update touched three files (+20 / -10) and brought the host guide in line with the current chain (PR #1165):

  • The guide now lists the live model roster - Qwen, Kimi, and MiniMax - instead of the Kimi-only description it carried before.
  • It records the current weight_scale_factor and penalty_start_epoch values for each of the three models. These two parameters control how much a model contributes to a validator's weight and the epoch at which under-performance starts to incur a penalty.
  • Host guidance that was written specifically for Kimi was generalized to per-model wording, so operators running any of the three models read instructions that apply to their setup.
  • The English and Chinese release announcements were corrected: MiniMax's penalty_start_epoch is 278 on the live chain, not the 271 written in the original ticket (GON-171).

The author verified the numbers against the live chain parameters endpoint (node3.gonka.ai/chain-api/.../inference/params) and built the docs site locally with mkdocs build --strict before merging.

Why it matters

Hosts size and configure their machines from these docs. When the guide describes one model but the chain runs three, an operator can misjudge which models they are eligible to serve and how their weight is calculated. Aligning the page with the governance-approved model set closes that gap.

The penalty-epoch correction matters for the same reason. A host reading 271 would expect MiniMax penalties to begin seven epochs earlier than they actually do. Matching the docs to the on-chain value of 278 means the release notes no longer contradict the chain about when penalties start.