IBM Selects NVIDIA Nemotron for Self-Hosted IBM Bob Deployment

IBM has announced the general availability of a self-hosted deployment option for IBM Bob, its agentic software development platform. The move is aimed at supporting sensitive and mission-critical workloads on enterprise-controlled infrastructure. To ensure optimal performance, IBM evaluated both open-weight and closed-weight models across four key criteria: hardware footprint, coding accuracy, latency, and openness.

The evaluation process involved benchmarking leading models from Anthropic, OpenAI, and Google against open-weight alternatives. IBM ultimately selected NVIDIA Nemotron and Poolside Laguna for their superior performance. NVIDIA Nemotron 3 Ultra was praised for its consistent output behavior, low latency, and reliable task completion. Meanwhile, Poolside Laguna S 2.1 stood out for its reasoning capabilities, fast response generation, and strong performance-to-footprint ratio.

IBM collaborated closely with NVIDIA to optimize the Nemotron model for IBM Bob’s agent harness. Through prompt tuning, sampling parameters, and reasoning settings, the team improved the model’s accuracy and efficiency. The resulting models were then tested by IBM Bob’s development and usage teams to ensure real-world applicability.

For customers, the self-hosted deployment offers straightforward installation and hosting on their own infrastructure. This ensures that code, IP, and context never leave the enterprise environment. The models are tuned specifically for IBM Bob’s harness, delivering fast and reliable results out of the box. Whether in regulated, air-gapped environments or self-managed deployments, IBM Bob self-hosted is designed to provide a seamless experience with support from partners like NVIDIA.