A new version of MLNode is on its way, built on an updated vLLM 0.15.1 inference engine. Better performance on newer GPUs and a path to adopt Kimi K2.5.
What it unlocks
New MLNode is under review and in stability testing now, with further refinements expected. It’s compatible with existing MLNode hosting Qwen235 and both versions can coexist in the network.
This upgrade improves how MLNode utilizes newer hardware. Latest-generation GPUs, including B200, demonstrate significantly better performance in both inference and PoC, with the node now able to take fuller advantage of their capabilities.
In addition, the new MLNode unlocks the path to adopt Kimi K2.5 on 8×H200 / 8×B200 setups.
How it works
The update follows the standard MLNode release process — no on-chain vote required. Each operator decides independently and on their own timeline. Different versions of MLNode already coexist in the network, and that remains the norm.
The ability to support Kimi (or any other new model) does not automatically change what is running in the network. Adopting a new model would still require an on-chain vote. However, releasing a compatible MLNode version is a necessary step before such a vote can take place.
For those planning to upgrade: full release notes and deployment instructions will be shared shortly. No action required at this stage.