What Is Post-Training in AI? Base Model to Product

Post-training in AI is everything a lab does to a model after the pretraining run finishes. Pretraining produces something that can continue text plausibly and little else. Post-training is the stage that turns that into a model which follows instructions, holds a conversation, calls tools, refuses certain requests and has a recognisable personality. Same weights to start with, very different product at the end.

If you have ever wondered why two models built on similar architectures and similar data behave nothing alike, post-training is usually the answer.

The one-line difference

Pretraining teaches a model about the world. Post-training teaches it how to behave.

Pretraining is the expensive, months-long pass over a very large corpus that produces a base model. A base model is capable but unhelpful: ask it a question and it may well continue with three more questions, because that is what a page of questions usually looks like. Post-training is the shorter, far cheaper stage that makes it answer.

What is actually inside post-training

It is a pipeline rather than one technique, and the exact mix differs per lab. The recurring stages:

• Supervised fine-tuning. Show the model thousands of examples of a good response to a prompt and train on them directly. This is what installs the basic instruction-following reflex.

• Preference optimisation. Show the model two candidate answers with a signal about which humans preferred, and push it toward the preferred one. RLHF is the best known version of this, and it depends on a reward model standing in for human graders at scale.

• Reinforcement learning on verifiable tasks. Where correctness can be checked automatically, maths with a known answer, code that either passes tests or does not, you can train against the checker instead of against human opinion. This is where most of the recent gains in reasoning and coding models come from.

• Safety and refusal tuning. Teaching the model which requests to decline and how. It is also where over-refusal creeps in, which is the mechanism behind models refusing perfectly safe questions.

• Tool and format training. Reliable JSON, consistent function-call syntax, knowing when to reach for a tool rather than guess.

Post-training is not the same as fine-tuning

This is the distinction most people get wrong, and it matters when you are reading a release note.

Fine-tuning is one technique. Post-training is the umbrella over all of the stages above, fine-tuning included. When a lab says a model was post-trained, they mean the whole pipeline that produced the assistant you talk to. When you fine-tune a model, you are doing one narrow adaptation on top of a model that has already been through somebody else's full post-training.

The practical consequence: your fine-tune inherits the lab's post-training decisions and cannot easily undo them. If the base assistant hedges, your fine-tune will tend to hedge. If it refuses a category of request, a few hundred of your examples usually will not change its mind. You are decorating a room, not rebuilding the house.

A worked example from this week

Cognition released SWE-2 on 10 September 2026 and described it as post-trained from Kimi K3, a 2.8-trillion-parameter model that had already been through extensive reinforcement learning for agentic coding. Cognition then ran its own post-training on top, producing three effort levels in a single training run.

That is post-training doing exactly what it says. Nobody trained a new frontier model. One lab took another lab's already coding-tuned base and spent its budget on the behaviour layer, and the result landed within a point of a frontier coding model on Cognition's own benchmark. The capability was largely in the base. The product was in the post-training.

Why this matters when you are just using models

Three things follow that are worth carrying around.

Model updates can change behaviour without changing capability. A point release that keeps the same base and adjusts post-training will feel different in ways benchmarks do not capture: more verbose, more cautious, differently formatted. That is the usual explanation when a model seems worse after an update despite scoring the same or higher.

Prompt tricks are post-training artefacts. The reason a phrasing that works beautifully on one model does nothing on another is that you are exploiting a habit installed during someone's post-training, not a property of language models. That is the real reason prompts do not transfer cleanly across models.

Open weights do not mean identical models. Two products built on the same open base can behave completely differently, because the post-training is where the differences live and it is rarely released alongside the weights.

Frequently asked questions

Is post-training the same as RLHF?

No. RLHF is one stage inside post-training, and an increasingly small one. Many current models lean more on reinforcement learning against automatic checkers than on human preference data, particularly for maths and code.

How long does post-training take compared to pretraining?

Far less, in both time and compute. Pretraining is the months-long run that dominates the budget. Post-training is measured in weeks and is cheap enough that labs iterate on it repeatedly between base-model generations, which is why you see point releases far more often than new base models.

Can I post-train a model myself?

You can run pieces of it. Supervised fine-tuning and lightweight adapter methods such as LoRA are within reach of a small team on rented GPUs. A full preference-optimisation pipeline with a trained reward model is a different order of effort and rarely the right call when retrieval or a longer context would solve the same problem.

Does post-training add new knowledge to a model?

Mostly not. Facts come from pretraining. Post-training shapes how the model uses and presents what it already knows, which is why it cannot fix a stale training cutoff.

For the wider picture of how these stages fit together, our overview of how AI models work walks the path from raw text to the assistant you actually talk to.