IBM says it has tuned NVIDIA Nemotron models for IBM Bob, its self-hosted AI coding assistant, giving organisations a way to run agentic software development on infrastructure they control. The practical draw is local deployment for sensitive or air-gapped environments, alongside IBM’s claim of up to four times lower latency on NVIDIA Blackwell GPUs than leading SaaS models accessed over the network.
Watch Desk analysis
What happened
IBM’s engineering team says it assessed open- and closed-weight models for hardware footprint, coding accuracy, latency and openness, testing them with Bob’s agent harness on software-engineering workflows of varying difficulty. It selected NVIDIA Nemotron 3 Ultra and Poolside Laguna S 2.1 for the self-hosted product.
IBM says its team worked with NVIDIA on prompt tuning, sampling settings, reasoning settings and changes to the agent harness. The company says Nemotron 3 Ultra followed tool schemas reliably, avoided unnecessary complexity in code and delivered up to four times lower latency on Blackwell GPUs than leading SaaS models. Those latency figures come from IBM’s internal benchmarks. IBM also says Laguna S 2.1 performed strongly relative to its size and hardware footprint.
Why it matters
For organisations that cannot send source code, intellectual property or context to an external service, running a coding agent on their own infrastructure may make deployment possible where a hosted assistant is not. The model is only part of that proposition: IBM’s evaluation included Bob’s agent harness, which affects how models perform when they use tools and complete real coding tasks.
The trade-off is that self-hosting puts hardware, installation and operation on the customer. IBM’s results are useful evidence of what its team found in its own product, but the reported speed advantage is not a general comparison across every model, task or deployment.
Our read
This is a concrete enterprise-AI development, not just another model name in a busy catalogue: IBM has paired a self-hosted coding product with selected models and says it has tuned the full agent setup. Teams considering it should ask for benchmark details on their own workloads, and compare the cost of running the hardware with the convenience of a hosted service. “Four times faster” is a promising headline; the workload and setup are the fine print that decide whether it matters.
What to watch
- Which IBM Bob self-hosted customers can access the Nemotron and Laguna options, and what hardware they need.
- Whether IBM publishes benchmark methods, workloads and comparisons behind its latency claims.
- How the models perform on customers’ own codebases and tool-using tasks.
Discussion spark: For sensitive codebases, is the control of self-hosting worth the hardware and maintenance burden, or should teams prefer a hosted coding assistant unless they have no choice?
Sources and evidence
- IBM chose NVIDIA Nemotron for IBM Bob self-hosted deployment – IBM (5 October 2026, 18:23 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.