IFM releases PAN 2 world-model weights and training code
MBZUAI’s PAN 2 pairs language planning with video simulation. Here is what is open, what hardware it needs and where its plans still fail.
vLLM
NVIDIA NemoClaw update narrows local vLLM access
NemoClaw 0.0.142 changes managed vLLM port bindings and agent onboarding. Existing containers need separate checks before the new defaults apply.
Buoyant launches Kinference for Kubernetes, with alpha limits
Kinference brings policy and cache aware routing to vLLM on Kubernetes. Its alpha still lacks rate limits, semantic classifiers and Anthropic style calls.