Skip to content
nashiaiguide

Getting good answers from a small model

How do you get good answers from a small AI model on your phone? Give it the material, ask for one thing at a time, and prefer rewriting to inventing.

By Updated 3 min read

Last verified: Nashi 1.0, October 2026

On this page

To get good answers from a small AI model, give it the material instead of asking it to remember, say exactly what you want in the first line, ask for one thing per message, and use it to rewrite rather than invent. Small models read well and know little. Play to that and they do good work.

What should I expect from a small AI model?

A model small enough to run on your phone, like the ones in Nashi, is a capable writer and a reasonable explainer. It is not a search engine, not a calculator, and not a legal department. Use it for what it is good at, and it is genuinely useful. Ask it for the population of a town in 1890 and it will guess with full confidence.

Why does a small model make things up?

Small models know less, but they read fine. Paste the email you are replying to, the paragraph you want rewritten, the notes you want summarised. Everything you give the model in the conversation is available to it; everything you don't, it will make up.

Nashi's models see about 3,000 words at once. If a conversation runs longer than that, the oldest turns fall out of view. For a long task, start a fresh conversation and paste in only what matters.

How should I phrase a prompt for a small model?

Small models follow clear instructions well and infer intent poorly. Compare:

"Here are my notes from the meeting."

with

"Turn these meeting notes into five bullet points for an email to my team. Keep names. Under 100 words."

The second gets what you asked for the first time. Length, format, audience and tone are the four things worth stating every time.

Can I ask for several things at once?

"Summarise this, then translate it to Spanish, then make it funnier" is three tasks, and a small model will usually do the first well and drift on the rest. Ask, read, ask again. The conversation keeps the context.

Should I ask it to write from scratch?

Ask the model to improve, shorten, restructure or change the tone of text you wrote, and it shines. Ask it to write a factual article from scratch, and you have to check every claim. That is true of every language model; it is simply more true of small ones.

When do I need a bigger model?

Download a larger model in Nashi, if your device shows one as a good fit (see choosing a model in Nashi), when you keep hitting one of these:

  • It loses the thread in a long conversation.
  • It ignores half of a detailed instruction.
  • You need code that runs, not code that looks right.
  • You need it to reason through several steps, not just recall.

And for a few things, use a cloud model and accept the trade: current events, obscure facts, long documents, anything where being wrong is expensive. Nashi is the place for everything you would rather not send anywhere.

What is it good for, day to day?

  • Drafting and rewriting emails, messages and posts.
  • Summarising notes, articles you paste in, and your own rambling.
  • Explaining a concept in plainer words, or at a different level.
  • Brainstorming names, titles, outlines and options.
  • Translating short passages between common languages.
  • Being a patient rubber duck at 2 a.m., with no server on the other end.

Frequently asked questions

Why does a small AI model make things up?

It knows far less than a large model, and it will still answer with confidence. Paste in the facts you want it to use, and check anything it states that you did not give it.

How much text can Nashi's models read at once?

About 3,000 words. Every model in Nashi has a 4,096-token context window. When a conversation runs longer, the oldest turns fall out of view, so start a fresh chat for a new long task.

What is a small on-device AI model good at?

Rewriting, shortening and changing the tone of your own text, summarising notes you paste in, explaining concepts, brainstorming, and translating short passages between common languages.

nashiaiguide

Choosing a model in Nashi

Which AI model should you run on your iPhone? Start with Gemma 3 1B, then go bigger if answers feel thin. How Nashi matches its nine models to your memory.

5 min read

aiprivacynashi

Why offline AI matters

Can an AI chatbot work offline? Nashi runs open models on your iPhone, even in airplane mode. What stays private, what you give up, and who it is for.

4 min read