Yes. IBM Bob’s self-hosted deployment became generally available on September 24, 2026. Its backend runs on a customer-operated Red Hat OpenShift cluster, either on premises or in the customer’s cloud account. But self-hosting the backend does not automatically keep model inference local: organizations can connect to supported external model services, or run supported models on their own infrastructure, including in an air-gapped environment.
What “self-hosted IBM Bob” means
IBM packages Bob’s backend as workloads for Red Hat OpenShift. The customer operates the cluster, while an operator manages installation, upgrades, and day-two operations. IBM says the backend includes identity, an inference gateway, audit logging, and usage metering, and can run in a namespace alongside other applications. Developers continue to use Bob IDE and Bob Shell; IBM says the underlying agent harness is the same as in its cloud service. IBM’s deployment overview describes the architecture and installation approach.
The key distinction is between where Bob’s backend runs and where a selected model processes requests. The backend can be on the customer’s cluster while model requests and accompanying code context are sent to an external service. To assess data handling, decide on the model route first, then establish where prompts and code context are processed under that configuration.
Where inference happens: the available routes
| Route | Where the backend runs | Where model inference happens | What to consider |
|---|---|---|---|
| Self-hosted backend with an external model service | Customer-operated OpenShift cluster | At the supported external model service connected through the customer’s cloud account or approved hybrid configuration | IBM says requests and included code context go to that cloud account under the customer’s existing agreements. Confirm the model’s eligibility and data handling with IBM and the service provider. |
| Fully self-hosted inference | Customer-operated OpenShift cluster | On customer infrastructure, using customer GPUs | IBM describes this route for networks without outbound connectivity, including air-gapped environments. Hardware sizing depends on vendor guidance, quantization, context length, and the number of developers served. |
IBM’s October 1, 2026, release post names NVIDIA Nemotron 3 Ultra and Poolside Laguna S 2.1 for fully self-hosted inference, and says a small guardrail model screens inputs and outputs on that route. Its September 30 announcement also lists Claude Sonnet 5.0, Claude Opus 4.8, Gemini 3.7 Flash, and OpenAI GPT 5.6 as external options for self-hosted and hybrid/private-SaaS configurations; it names NVIDIA Nemotron and Poolside Laguna among self-hosted options. Model names and eligibility can change, so ask IBM for the current support matrix before selecting an architecture. IBM’s September 30 announcement summarizes deployment configurations and named models.
#1 Best Overall
IBM says customers source and provide access to a supported model at general availability and describes a bring-your-own-license approach for eligible existing model licenses. That does not establish that every model works in every configuration or is included in the Bob license; confirm model access and licensing for the intended deployment.
What the OpenShift installation involves
IBM describes a two-stage installation using bobctl. The first stage installs cluster-wide resources, permissions, and the operator; it requires cluster-admin access. The second installs Bob into a namespace and requires namespace-level access. IBM says a single-node OpenShift deployment is enough for a proof of concept. For an air-gapped installation, IBM says bobctl can mirror images to a private registry, including transfer across the network gap. These are IBM’s published installation details, not an independent assessment of deployment effort or production readiness.
- Prepare the OpenShift environment. Confirm cluster access and determine who can approve the cluster-wide resources and operator installation.
- Install the operator and cluster resources. Use
bobctlwith cluster-admin access for this first stage. - Install Bob in a namespace. Complete the second stage with namespace-level access.
- Connect identity and model access. IBM says the deployment includes Keycloak and can connect to LDAP or Active Directory. Smaller installations and proofs of concept can manage users directly in the identity service. Configure access to the supported model chosen for the deployment.
- Plan lifecycle ownership and production capacity. Establish responsibility for operator-managed upgrades and day-two operations. IBM does not publish detailed production sizing in the cited deployment post; for local inference, sizing depends on workload and vendor guidance.
When this deployment may fit
- Organizations that need backend control: The customer operates the OpenShift cluster and retains the backend components there.
- Teams with strict network boundaries: IBM describes fully self-hosted inference for air-gapped environments, subject to supported models and suitable customer infrastructure.
- Teams that can use external models under their own arrangements: A hybrid route keeps the Bob backend on the customer cluster but sends requests and code context to the configured external service.
- Modernization teams: IBM offers optional Premium Packages for Java Modernization, IBM i, and IBM Z. These are separate licensed extensions with applicable package and deployment requirements. IBM’s IBM Z announcement describes capabilities such as deterministic application understanding, static code analytics, Z-specific context, dependency and business-rule discovery, and modernization workflows including Application View and zContext. IBM’s IBM Z package announcement provides its stated scope.
IBM positions self-hosting around control, security, governance, and sovereignty. The deployment description alone does not establish compliance with a particular regulation or guarantee a productivity, cost, or security outcome; those depend on the complete configuration and how it is operated.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to evaluate it before adoption
IBM describes the offer as sales-led and directs prospective customers to an IBM representative, Business Partner, or Contact Sales for a demonstration or proof of concept. Use that process to settle the deployment and commercial details that are not specified in the public product posts.
Quick Recap
- Which models are currently supported for the intended self-hosted or hybrid configuration?
- Where are prompts and code context processed for each model route, and what agreements govern that processing?
- What OpenShift prerequisites and production sizing apply to the organization’s workload?
- Who will own cluster-wide installation approvals, identity integration, upgrades, and ongoing operations?
- Which optional modernization packages are needed, and what licensing and deployment requirements apply?
- What pricing, service-level commitments, and contract terms apply? The cited public announcements do not state them.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
