IBM ties Bob’s self-hosted model choices to the agent harness, not benchmark scores alone

IBM has explained why its self-hosted Bob coding agent supports NVIDIA Nemotron 3 Ultra and Poolside Laguna S 2.1, describing a selection process that evaluated the models inside Bob’s agent harness rather than relying only on standalone model benchmarks.

The distinction matters for teams planning Bob on Red Hat OpenShift. Model selection becomes part of the platform design: operators must balance coding behavior and latency against GPU capacity, while the harness determines whether a model reliably follows tool schemas and completes multi-step work.

Four tests, one system boundary

In its engineering account, IBM says it evaluated open- and closed-weight models on hardware footprint, coding accuracy, latency and openness. The company also tested the models together with Bob’s harness across enterprise coding tasks, including tool calls and terminal operations.

IBM says early Nemotron testing showed strong reasoning but excessive verbosity, which increased the number of turns needed to finish work. IBM and NVIDIA then adjusted prompts, sampling parameters and reasoning settings while fixing the surrounding harness. That account makes the supported model list a product-level integration result, not a claim that either model is universally best.

IBM attributes Nemotron 3 Ultra’s selection to consistent tool-schema behavior, relatively direct solutions and lower latency. The company reports latency up to about four times lower on NVIDIA Blackwell GPUs than leading software-as-a-service models reached over a network, but labels that figure an internal benchmark and does not publish enough methodology in the post for independent comparison. IBM says Laguna S 2.1 offered a favorable performance-to-footprint ratio and fast response generation for more resource-constrained installations.

The OpenShift consequence

IBM’s general-availability announcement lists Nemotron and Laguna as the supported self-hosted choices. It also allows supported external services in hybrid configurations, so platform teams can decide where inference occurs according to workload controls.

Bob itself runs as a tenant workload on a customer-managed OpenShift cluster. IBM’s current system requirements support OpenShift Container Platform 4.20 through 4.22 on amd64 workers and give a headroom-adjusted Bob Core footprint of roughly 36.5 vCPU, 53.4 GiB of memory and 50 GiB of persistent storage. Those figures cover the Bob application stack, not the separate accelerator footprint needed to host Nemotron or Laguna.

That leaves a practical planning boundary. The Bob cluster requirements describe the agent service, storage and supporting components; the model choice determines an additional inference estate. Teams considering an air-gapped or fully self-managed deployment should therefore size and validate both halves together, including model-serving throughput under the actual Bob harness rather than extrapolating from generic coding leaderboards.