**Option A is correct.** A base model is the raw output of pre-training, where the only training objective was next-token prediction over a giant corpus of text. Hand it a question and it often does not answer; it continues the text it sees, sometimes by writing what looks like the rest of a forum post or a sibling document.
An instruction-tuned model has gone through at least one additional training stage on top of the base. Supervised fine-tuning (SFT) on prompt and response pairs teaches the model to actually answer when asked a question, follow a chat template, and respond in a helpful conversational style. A preference-tuning pass like RLHF or DPO usually follows, which shapes the model to prefer helpful responses and to decline harmful ones.
The user-visible difference is the combination of these two behaviours. The instruct model answers your question rather than continuing your sentence, and it refuses categories of unsafe requests rather than complying with them.
The wrong options locate the difference in the wrong place. The weights are genuinely different (so B is wrong). Size and training stage are separate concerns (so C is wrong). Safety is learned into the weights, not bolted on at decode time (so D is wrong).
**The right answer is A**, and the precise reason matters. A base model is the output of pre-training, where the objective is next-token prediction over a large heterogeneous corpus. The model learns rich representations of language but is not conditioned to answer questions or follow chat-style prompts. Asked 'What is photosynthesis?', a base model might continue with '... is a process by which plants...' which looks like an answer or with 'this question appears on many biology exams', which is what a sibling document would say. The behaviour is not adversarial; the model is just doing what it was trained for.
An instruction-tuned model has had at least one additional post-training stage. The first is supervised fine-tuning on instruction-response pairs, which teaches the model to recognise a chat template, treat the user turn as a request, and emit a response. The second, usually a preference-optimisation pass like RLHF or DPO, shapes the model to prefer helpful, harmless, and honest responses according to a preference signal. After this stage, the model follows chat-style asks reliably and refuses categories of unsafe requests.
The weights are genuinely different. The same architecture and starting point produce different parameter values after the post-training passes, which is why a chat template alone cannot turn a base model into a chat model. The distractors get this wrong: B claims the weights are identical, C confuses size with training stage, and D invents a runtime safety classifier when safety in instruct models is learned into the weights through preference tuning. The headline distinction is two behaviours that emerge from post-training: instruction-following and trained refusals.
Quick reference
TL;DR
Base = raw next-token predictor over pre-training data. Instruct = base plus SFT on prompt-response pairs plus usually preference tuning. The difference lives in the weights, not in templates or classifiers.
Picture two musicians who graduated from the same conservatory. The base musician knows how to play every scale and copy any style they have heard, but if you ask them to perform at your wedding they might just keep practising scales because nobody told them that a request means perform. The instruction-tuned musician went through extra coaching after the conservatory: a teacher sat with them and walked through 'when someone asks for a wedding song, this is how you respond'. They also learned 'when someone asks you to do something harmful, decline politely'. Both musicians have the same instrument and the same training in music theory. The instruction-tuned one has additional habits baked in through that follow-up coaching.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Building a deeply custom assistant whose persona, format, or safety profile diverges from the vendor's instruct release Starting from the base model gives a clean canvas; the vendor's instruct post-training would have to be partly undone before your fine-tune can take hold.
Quick deployment for general chat use with no custom training Pick the instruct variant; it already has the chat conditioning and safety alignment the vendor intended for ship as-is use.
Comparing models on pretraining benchmarks like perplexity on a held-out corpus Use base variants for these comparisons; instruct post-training shifts the output distribution in ways that confound raw pretraining evaluations.
Quick scan · 2–4 min
Everything important, quickly.
4 min: pre-training objective + SFT and preference tuning stages + two emergent behaviours (chat-following and refusals) + why the weights actually differ + when to start from base versus instruct.
Real products, models, and research that use this idea.
1Llama 4 ships as both Llama 4 Maverick (base) and Llama 4 Maverick Instruct (post-trained); they share architecture and parameter count and differ in the post-training passes Meta applied.
2Qwen 3.5 has the same base/instruct split, with the instruct variant explicitly intended for chat use and the base recommended as a starting point for custom fine-tunes.
3Hugging Face TRL's SFTTrainer plus DPOTrainer is the canonical open-source recipe for taking a base model and producing an instruct variant.
4Anthropic's Claude Opus 4.7 and OpenAI's GPT-5.5 are not released as base models at all; only the post-trained variants are exposed via API, reflecting a deliberate product decision to ship only the instruct-style model.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhy does post-training change behaviour so much when it touches relatively few parameters by tokens-seen?
Sign in or go Pro to see the approach for this follow-up.
QWhat is lost when you fine-tune an instruct model further versus fine-tuning the base?
Sign in or go Pro to see the approach for this follow-up.
QHow does preference tuning produce refusal behaviour without an explicit refusal classifier?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Thinking the difference is just a chat template or a runtime safety filter. Post-training genuinely updates the weights; the same parameters produce different behaviour because they were optimised for a different objective.