Instruction-tuned model The version of a model that answers, rather than just continues.
Short answer. An instruction-tuned model is a language model that has had extra training so that it follows requests and answers questions, rather than only continuing the text it is given. A base model, fresh from pre-training, completes sentences: ask it a question and it may produce more questions. Instruction tuning, usually followed by preference tuning, turns it into an assistant. In the InstructGPT study, people preferred outputs from a 1.3B instruction-tuned model to those of the 175B base GPT-3, which has 100 times more parameters. Names show the difference: Llama-3.1-8B is a base model, Llama-3.1-8B-Instruct is tuned, and Gemma uses -pt and -it.
01
How it works
Pre-training teaches a model to continue text. Hugging Face’s documentation describes the next step: base models are fine-tuned for chat by training them on data formatted as sequences of messages. Instruction tuning uses pairs of an instruction and a good response. The FLAN study instruction-tuned a 137B model on over 60 tasks and found it beat zero-shot GPT-3 175B on 20 of 25 tasks. Preference tuning then teaches the model which of several answers people prefer. Meta describes the Llama 3.1 instruction-tuned models as using supervised fine-tuning and reinforcement learning with human feedback.
The LIMA study suggests why this works with little data: almost all of a model’s knowledge is learned in pre-training, and a small amount of tuning data teaches it how to present that knowledge.
02
How to tell them apart
Naming conventions differ, so read the model page rather than guess. The base and tuned versions below are all published in public repositories.
Family
Base model
Instruction-tuned model
Note
Llama 3.1
Llama-3.1-8B
Llama-3.1-8B-Instruct
Suffix -Instruct marks the tuned version
Gemma 3
gemma-3-27b-pt
gemma-3-27b-it
pt for pre-trained, it for instruction-tuned
Qwen3
Qwen3-8B-Base
Qwen3-8B
The model card of the Base version says it is pre-trained without instruction tuning
03
Which one to use
Assistants, chat, document questions and agents: the instruction-tuned version.
Text completion experiments, building your own tuned variant from scratch, or research: the base version, which has not been shaped by a chat format.
Check whether a tuned model also has a reasoning mode. Qwen3-32B, for example, can think before answering and can have that switched off.
If a model ignores your instructions and rambles on, you have probably loaded a base version.
04
Why it matters for an office
Downloading the wrong variant is a common cause of poor results that gets blamed on the model. A base model asked to “summarise this contract” may continue the contract instead. Instruction-tuned models also ship with a chat template, a fixed format for system, user and assistant messages, which the software must apply correctly. For office use you almost always want the instruction-tuned version, then test it on your documents. Any tuning is also a place where behaviour can change between versions, so record the exact model name and version in your records.
05
Questions people ask
Is an instruct model the same as a chat model?
Usually they are used for the same purpose, an assistant that answers. Names vary by maker: -Instruct, -it or -chat. The model card says which kind of tuning was applied.
Can I use a base model as an assistant?
Only with careful prompting, such as giving it the start of an answer to continue. Instruction-tuned models are far easier to use for questions and tasks.
Does instruction tuning add knowledge?
Mostly not. The LIMA study concluded that almost all knowledge is learned during pre-training and tuning teaches the model how to respond.
What is preference tuning?
A further stage in which people, or automated judges, compare pairs of answers and the model is trained to prefer the better ones. InstructGPT used reinforcement learning from human feedback for this.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
