The current frontier of AI systems is based on two model types(they are almost identical under the hood, but their behavior is notably different in practice, hence the distinction)

Pre-trained models, also known as ‘non-reasoning models’

These are the famous ‘Large Language Models’, or LLMs, gigantic AI models trained on as much as data as possible, reaching double digits of trillions of words (for reference, Lama 3.1 405B was trained on 15 trillion tokens ~ 11-12.5 trillion words, and DeepSeek v3 14.8 trillion tokens, in the same rage).

Examples include GPT-4 (OpenAI), Opus (Anthropic), Gemini 2.0 (Google), or Grok-2 & 3 (xAI, the latter of which remains unreleased).

Their biggest characteristic is how they approach a response: they are fast thinkers. Upon receiving the user’s request, they immediately commit to a response with no hesitation. Think about them as ‘intuition machines’ as if you always responded to questions using your immediate intuition.

If you’re one for analogies, they would be similar to how Homer or Peter Griffin respond; not much filtering between what the brain first thinks and what gets spit out.

Reasoning models, also known as Large Reasoner Models, or LRMs

The talk of the town right now, they behave slightly differently. Instead of simply committing to the first thing that comes to mind, they take a multi-stop approach to answering, slower, more thoughtful thinking, just like you would when receiving a complex task to solve.

Why do we want this?

As Noam Brown, OpenAI’s Reasoning Lead, puts it, “Some problems benefit from you thinking for longer on them.” This means reasoning models don’t immediately commit to answering and instead will reflect, iterate, backtrack, and search for alternatives if the current thought does not meet user’s demands until converging into a response. Think about this process as the one you would when trying to solve a complex math problem.

Naturally, these models are conceived for solving complex problems, but do not represent any advantage over LLMs to solve problems that do not require long thinking, like answering knowledge-based answers such as ‘What’s Poland’s capital.’