NVIDIANVIDIA-Certified Professional: Agentic AI2026-100 Prerequisites
Choose models by task and training
Completed by the knowledge check
Learning objective
Distinguish foundation, instruction-tuned, and reasoning LLMs.
Core idea
Separate what a model learned during training from how you ask it to behave during inference. A foundation large language model (LLM) provides broadly reusable language capabilities. Instruction tuning adapts a pretrained model to respond to instructions. Reasoning models support additional inference-time reasoning before producing an answer.
These categories overlap: an instruction-tuned reasoning model still builds on a pretrained foundation. They describe training and behavior, rather than three interchangeable deployment products. For an agent, the useful question is which behavior satisfies the task within its latency and quality constraints.
How it works
A base language model learns patterns from a training corpus and predicts continuations. That capability provides a starting point for adaptation; it does not establish that a checkpoint follows a particular assistant’s instruction format reliably.
Supervised fine-tuning (SFT) changes model parameters using examples of inputs and desired outputs. Instruction tuning teaches the model to follow the supplied instructions and respond in the demonstrated style. Writing a system prompt at request time supplies input; it does not perform SFT or change the weights.
NIM’s NVIDIA-Nemotron-3.5-Super-VL-120B-A12B deployment enables reasoning by default. Set chat_template_kwargs.enable_thinking: false to disable it per request, or use reasoning_effort: low and medium select shorter reasoning; high selects full reasoning. Reasoning tokens count toward max_tokens, so leave room for the final answer. Do not use this model’s thinking_token_budget to cap reasoning: reaching it leaves content empty; use the documented effort or thinking controls instead.
Response fields also depend on the model: Nemotron returns reasoning in choices[].message.reasoning, while NIM’s GLM-5.2 page documents choices[].message.reasoning_content. Neither the control nor the response field is universal.
You must budget generation for reasoning and the final answer. More generated reasoning consumes inference work and time; evaluate whether the resulting task improvement earns that cost. See 3.4.2 for model size and reasoning-budget selection, and 5.2.1 for reasoning modes within token budgets.
Key distinctions and trade-offs
Training versus prompting: Fine-tuning changes weights. A prompt changes the input to an inference request. A strong instruction prompt does not turn a base checkpoint into an instruction-tuned one.
Reasoning versus action: Reasoning produces model output; a tool executor changes or queries the environment. A reasoning-capable model does not, by that label alone, implement an agent loop. See 0.4.
Capability versus suitability: A model that solves difficult planning questions can still be a poor choice for a brief extraction task with a tight response target. Compare completed task quality and latency, rather than selecting by the category name.
Architecture / decision scenario
A manufacturer’s employee assistant must classify maintenance tickets within a two-second response target. In an illustrative internal test, an instruction-tuned candidate achieves 96% acceptance at 1.2 seconds; a reasoning-enabled candidate achieves 97% at 4.8 seconds. A base checkpoint requires additional adaptation before it follows the ticket format consistently.
The tempting proposal is reasoning for every request because its acceptance rate is higher, but that path misses the target. The assistant owner also tests a shared reasoning-capable deployment with enable_thinking: false for classification and thinking enabled for investigations. This leaves one deployment to operate, but its classification latency and accuracy still need measurement. A smaller dedicated instruction model adds a deployment to operate and must earn that cost against the shared option’s measured results. The owner keeps classification on the dedicated candidate unless the shared option meets two seconds and at least 96% acceptance with thinking off. The model team owns base-model adaptation; changing a prompt is a different intervention.
Common exam traps
- Treating foundation, instruction-tuned, and reasoning as mutually exclusive architectures.
- Calling a request-time prompt a weight update.
- Assuming longer reasoning guarantees a better answer or successful tool execution.
- Applying one model’s thinking control to every NIM deployment.
What to remember
Identify the training stage, the documented inference behavior, and the task’s operating constraints separately. Choose instruction following or reasoning from measured task results, and verify the model-specific request interface.
Knowledge check
A team has a base checkpoint that does not follow its ticket-classification format consistently. A colleague proposes writing a detailed system prompt with the format and worked examples, and then calling the result an instruction-tuned model. Which statement about the proposal is accurate?
An agent calls NIM’s NVIDIA-Nemotron-3.5-Super-VL-120B-A12B deployment, which reasons by default. For a routing step you want the model to keep reasoning, but in a shorter mode. You leave max_tokens high enough for reasoning and a final answer in content. Which request change fits the model’s documented controls?
thinking_token_budget to a small value for the routing requestchat_template_kwargs.enable_thinking to false for the routing requestreasoning_effort to low for the routing requestmax_tokens for the routing request to shorten its reasoninglow and medium effort select a shorter reasoning mode, and with max_tokens left high enough for both reasoning and the answer, the response ends in content . Reaching thinking_token_budget leaves content empty, enable_thinking: false removes reasoning rather than shortening it, and a lower max_tokens shrinks the allowance that reasoning and the answer spend together, so the response can end inside the reasoning.A client written against NIM’s GLM-5.2 deployment logs reasoning from choices[].message.reasoning_content. The team points it at NIM’s NVIDIA-Nemotron-3.5-Super-VL-120B-A12B deployment with reasoning on: final answers arrive in content, but the reasoning log stays empty. What is the most likely cause?
reasoning_effort is highchoices[].message.reasoningreasoning_content , while Nemotron returns reasoning in reasoning , so a client reading the GLM field finds nothing. Nemotron reasons by default, and its low and medium efforts still reason, only more briefly; neither a control nor a field carries over from one model’s page to another.You already operate a reasoning-capable NIM deployment for investigations. A new extraction step must answer within 1.5 seconds with at least 95% acceptance, and a dedicated instruction-tuned model would be one more deployment to operate. What should decide whether extraction runs on the shared deployment?