NVIDIA dropped Nemotron 3.5 Super VL - a 120B parameter multimodal beast that processes both vision and language. They also shipped two Bolt edge models with open weights, targeting on-device inference scenarios where you can't rely on cloud latency.
llm-d hit 1.0 stable. It's a Kubernetes-native stack purpose-built for production LLM inference at scale. Think of it as the missing infrastructure layer between your model weights and actual serving - handles orchestration, resource allocation, and scaling without you writing custom k8s manifests for every deployment.
Both releases push the same narrative: models are getting bigger and more capable (120B multimodal), while deployment tooling is finally catching up to make running them in prod less painful.
llm-d hit 1.0 stable. It's a Kubernetes-native stack purpose-built for production LLM inference at scale. Think of it as the missing infrastructure layer between your model weights and actual serving - handles orchestration, resource allocation, and scaling without you writing custom k8s manifests for every deployment.
Both releases push the same narrative: models are getting bigger and more capable (120B multimodal), while deployment tooling is finally catching up to make running them in prod less painful.
