OpenShift AI workflow navigator narrows model choices, but operators still own the proof

Red Hat’s workflow navigator developer preview aims at a real source of AI cost: selecting a model and sizing a deployment without workload-specific evidence. Its useful boundary is narrower than the phrase “production-ready Kubernetes configurations” in the announcement may suggest.

What goes into the ranking

The user supplies a task and constraints such as concurrency and a latency target. Workflow navigator then checks the model registry, available cluster GPU hardware and benchmark data for the use case. Its planner filters models against hardware capacity and ranks candidates by relevant benchmarks.

That process can expose a cheaper candidate before GPU hours are committed. Red Hat’s example compares a 32-billion-parameter coding model on two GPUs with a 70-billion-parameter model on four GPUs when their code-benchmark results are similar.

The deployment side estimates GPU memory across quantization and tensor-parallel configurations. It can generate a KServe InferenceService, resource requests, autoscaling rules and observability configuration sized against the cluster inventory.

What the output does not guarantee

The article says the GPU recommender compares hardware performance without running actual benchmarks. Its result is therefore a plan assembled from registry data, known capacity and prior benchmark evidence—not a measurement of the proposed application under its real prompts, output lengths, arrival patterns or service dependencies.

Generated YAML can be syntactically complete while still encoding assumptions that do not hold in production. A memory estimate does not prove that tail latency meets an objective. An autoscaling rule does not prove that cold starts, queueing and quota pressure behave acceptably. A model’s score on a coding or RAG benchmark does not establish application quality for a private dataset.

Checks before deployment

Treat the ranked list as a shortlist. Benchmark the leading configurations with representative prompts, concurrency, context lengths and generation lengths on the target hardware. Measure p50 and tail latency, throughput, accelerator memory, cold-start behavior and failure recovery.

Review the generated KServe object like any other production change: pin model and runtime artifacts, verify namespaces and service accounts, set requests and limits, test autoscaling bounds, and connect the emitted observability settings to alerts that an operator will actually own. Run quality and safety evaluations against the intended data before accepting a lower-cost model.

Workflow navigator is available as a developer preview in OpenShift AI and exposes its workflow through MCP-compatible clients including Claude Code, Cursor and Gemini CLI. That makes it easy to place in an engineering loop. It does not change the final control point: the deployment should advance only after measured evidence confirms the navigator’s assumptions.