The Gonka team has released MLNode 3.0.12-post6, now set as the default version in the main branch. The update ships three Docker image variants covering standard, Blackwell, and Blackwell SM120 GPU architectures.
What Changed
The release addresses an edge case near Proof of Compute (PoC) start that could previously cause degraded performance under specific conditions, including scenarios involving reduced model weight. Several miners had already been running this version across recent epochs before it became the official default, providing real-world validation of its stability improvements.
The underlying vLLM engine moves from v0.9.1-pocv2-post5 to v0.9.1-pocv2-post6. The full diff is available on GitHub.
Available Images
| Variant | Image |
|---|---|
| Standard | ghcr.io/gonka-ai/mlnode:3.0.12-post6 |
| Blackwell | ghcr.io/gonka-ai/mlnode:3.0.12-post6-blackwell |
| Blackwell SM120 | ghcr.io/gonka-ai/mlnode:3.0.12-post6-blackwell-sm120 |
Upgrade Path
The update is fully backward-compatible with previous versions. Node operators can upgrade by pulling the new image:
docker pull ghcr.io/gonka-ai/mlnode:3.0.12-post6
The commit setting this as default landed via PR #990, updating all Dockerfiles from post4 to post6 base images.
Why It Matters
PoC stability near epoch boundaries is critical for validator uptime and reward consistency. Nodes operating close to PoC start were occasionally experiencing performance degradation that could affect their validation scores. This fix eliminates that edge case, meaning fewer missed validations and more predictable behavior for the entire network.