IBM Bob can now run on infrastructure that a customer controls, from on-premises servers and private clouds to sovereign clouds and air-gapped networks. The self-hosted option is meant for organizations that cannot send source code or regulated data to an outside AI service. IBM names financial services, government, healthcare and critical infrastructure as the sectors it has in mind.
IBM calls the self-hosted deployment generally available. At launch, customers supply the model IBM Bob uses, and the supported models in IBM's list fall into two groups. NVIDIA Nemotron and Poolside Laguna run on infrastructure the customer installs and manages. Claude Sonnet 5.0, Claude Opus 4.8, Gemini 3.7 Flash and OpenAI GPT 5.6 Sol connect through an approved private SaaS route to an outside model service. Customers holding eligible licenses for models on that list can reuse them under a bring-your-own-license arrangement.
Model placement decides where the code goes. IBM says a self-hosted model configuration lets code, development context and build artifacts stay inside the customer environment. On the hybrid route, the chosen model and configuration set where inference runs and how that material is processed. The sovereignty claim for IBM Bob therefore holds in full only for some model choices, a limit IBM's own wording acknowledges.
The self-hosted build keeps the core IBM Bob feature set, including the IDE experience, BobShell, parallel tool calling, an agent harness, and skills and modes. Optional IBM Bob Premium Packages add dedicated tooling for Java modernization, IBM i and IBM Z, each subject to its own licensing and deployment requirements. IBM sketches a bank that runs inference on-premises for core banking work while pointing a less restricted project at an approved external model service.
IBM backs the argument with its own research. A June report from its Institute for Business Value found that 68% of surveyed executives call meeting data residency and sovereignty requirements across geographies challenging. Neel Sundaresan, IBM's general manager of AI and automation, said organizations need "AI that operates inside environments they have control over."
IBM's announcement lists no pricing or hardware requirements for the self-hosted setup and points buyers to a demo. The company plans to widen the supported model list and add multi-model routing in later releases, with no dates attached, and its disclaimer says such plans can change. Until then, IBM names just two models that keep everything on customer hardware.
