Text, task by task
Pick your computer to see, task by task, whether a local model does the job or a cloud one like Claude, ChatGPT or Grok is better.
A PC with an RTX 4070 runs models up to about Qwen3 8B well. A local model can do 6 of these 11 tasks.
- Summarising articles and emailsLocalSmall models summarise well, and it costs nothing to run all day.
- Rewriting and fixing toneLocalA routine job that small models handle well.
- Sorting, tagging and pulling out detailsLocalRepetitive work across many files suits a model that is free to run.
- Anything with health, legal or client detailsLocalNothing you type leaves your computer.
- TranslatingEitherGood for common languages; weaker for rare ones. Stronger across more languages.
- Reading screenshots and photosEitherSome small models, like Gemma 3, can read images. Better at small text, charts and handwriting.
- Quick coding questionsCloudThis needs a mid-size model, and this machine runs small ones well. Small local models make more mistakes here; cloud models make fewer.
- Drafting and brainstormingCloudThis needs a mid-size model, and this machine runs small ones well. Small local models write flatter prose; cloud models have more range.
- Hard analysis, maths and reasoningCloudThe biggest cloud models are clearly ahead on multi-step problems.
- Working across a whole codebaseCloudLong, multi-step work over many files is where cloud models pull furthest ahead.
- Research on recent eventsCloudLocal models only know what they learned in training, and can’t search the web on their own.
A Mac mini M4 with 16 GB runs models up to about Qwen3 8B well. A local model can do 6 of these 11 tasks.
- Summarising articles and emailsLocalSmall models summarise well, and it costs nothing to run all day.
- Rewriting and fixing toneLocalA routine job that small models handle well.
- Sorting, tagging and pulling out detailsLocalRepetitive work across many files suits a model that is free to run.
- Anything with health, legal or client detailsLocalNothing you type leaves your computer.
- TranslatingEitherGood for common languages; weaker for rare ones. Stronger across more languages.
- Reading screenshots and photosEitherSome small models, like Gemma 3, can read images. Better at small text, charts and handwriting.
- Quick coding questionsCloudThis needs a mid-size model, and this machine runs small ones well. Small local models make more mistakes here; cloud models make fewer.
- Drafting and brainstormingCloudThis needs a mid-size model, and this machine runs small ones well. Small local models write flatter prose; cloud models have more range.
- Hard analysis, maths and reasoningCloudThe biggest cloud models are clearly ahead on multi-step problems.
- Working across a whole codebaseCloudLong, multi-step work over many files is where cloud models pull furthest ahead.
- Research on recent eventsCloudLocal models only know what they learned in training, and can’t search the web on their own.
- Estimate: Calculated from a stated rule of thumb. Real results vary.Which tasks suit which models: our judgement, from using themChecked: not yet verified
- Estimate: Calculated from a stated rule of thumb. Real results vary.What your machine runs well: the checker’s estimateChecked: not yet verified
Our judgement, not benchmark scores. “Either” means local does the job and cloud does it a little better.
Pictures, video and voice
Other models transcribe speech, read text aloud, or make pictures, video and music. They’re separate downloads, usually in separate apps, and they work your computer differently.
Text is limited by how fast your machine can Memory speedHow many gigabytes a second the chip can read from memory. The biggest single factor in how fast a model writes.Read the full entry. Pictures and video are limited by raw Computing powerHow much arithmetic a chip can do each second. Pictures and video need lots of it.Read the full entry: a Diffusion modelThe kind of model that makes pictures and video, by turning random noise into an image step by step.Read the full entry repaints the whole image from noise dozens of times. That’s where NVIDIA cards pull furthest ahead of Macs.
Turning speech into text
LocalLight on your machine
Transcribing meetings, interviews and voice notes, or dictating instead of typing.
Small, fast, accurate and private. One of the best reasons to run models at home.
- Models
- Whisper (OpenAI), Parakeet (NVIDIA)
- Apps
- MacWhisper and Superwhisper on Mac; Whisper-based apps on Windows
- Needs Estimate: Calculated from a stated rule of thumb. Real results vary.
- Any recent Mac, or a PC with a graphics card of 4 GB or more.
- On a Mac
- Works very well, often faster than the recording plays.
- On a PC
- Works very well, and very fast on an NVIDIA card.
Reading text aloud
EitherLight on your machine
Turning articles, notes or scripts into natural-sounding speech.
Local voices are good and free. Cloud voices are more expressive and easier to set up.
- Models
- Kokoro, Chatterbox
- Apps
- Mostly through tools for tinkerers, so far
- Needs Estimate: Calculated from a stated rule of thumb. Real results vary.
- Almost any computer. The smallest good voices are under 1 GB.
- On a Mac
- Works well.
- On a PC
- Works well.
Making pictures from words
EitherHeavy on your machine
Illustrations, concept art, mock-ups and photo edits, from a written description.
Local is free and private, with no limits on what you make. Cloud tools are simpler, and better at following long descriptions.
- Models
- FLUX (Black Forest Labs), Qwen-Image (Alibaba), Stable Diffusion (Stability AI)
- Apps
- Draw Things on Mac; ComfyUI on both
- Needs Estimate: Calculated from a stated rule of thumb. Real results vary.
- A graphics card with 12 GB or more, or a Mac with 32 GB. Smaller works, with smaller models.
- On a Mac
- Works, but expect tens of seconds a picture where a good card takes a few.
- On a PC
- The easier path. An NVIDIA card with 12 GB or more is comfortable.
Making video
CloudVery heavy on your machine
Short clips from a description or a still picture.
Unless you have a big NVIDIA card, cloud video tools are far faster and better.
- Models
- Wan (Alibaba), LTX-Video (Lightricks)
- Apps
- ComfyUI; Draw Things on Mac
- Needs Estimate: Calculated from a stated rule of thumb. Real results vary.
- A graphics card with 24 GB for good results. Smaller models run on 12 GB.
- On a Mac
- Possible, but very slow: a few seconds of video can take over an hour.
- On a PC
- Minutes per clip on a big NVIDIA card.
Making music
CloudHeavy on your machine
Songs or backing tracks from a description, with or without lyrics.
Open music models are early. Cloud services are well ahead on quality and ease.
- Models
- ACE-Step, YuE
- Apps
- Mostly through tools for tinkerers, so far
- Needs Estimate: Calculated from a stated rule of thumb. Real results vary.
- A graphics card with 8 to 16 GB, or a Mac with 16 GB or more.
- On a Mac
- Works, slowly.
- On a PC
- Works.
- Estimate: Calculated from a stated rule of thumb. Real results vary.Memory needs: published guides for each model, not our measurementsChecked: not yet verified
- Unverified: A stand-in value we have not confirmed yet.Mac video speed: one reported test, Wan 2.2 on an M1 MaxChecked: not yet verified
Side by side
| Local | Cloud | |
|---|---|---|
| Privacy | What you type stays on your computer. | Sent to the provider and handled under its terms. |
| Quality | Good for everyday work; behind the best cloud models on hard problems. | The strongest models available. |
| Cost | Free to run once you have the hardware, apart from electricity. | Free tiers with limits, then a subscription or pay as you go. |
| Speed | Depends on your machine; often faster than you can read. | Fast, but depends on your connection and how busy the service is. |
| Works offline | Yes. | No. |
| Limits | None, beyond your machine. | Usage caps, especially on free plans. |
| Knows recent events | Only up to when it was trained. | Often, through built-in web search. |
| Setup | About 30 minutes the first time. | Sign up and start. |
Privacy, plainly
When a model runs on your computer, your messages never leave it. No account, no server, no terms to change later. That’s the main reason people run models at home.
Cloud services handle your messages under terms that vary by provider and plan. Business plans often promise not to train on your data; free plans may not. If it matters, read the terms for your plan.
What local models can’t do yet
- Match the best cloud models on hard problems. The models that fit at home are smaller, and it shows on long reasoning, tricky code and subtle writing.
- Know about recent events. It only knows its training data, unless an app connects it to search.
- Hold very long conversations cheaply. Long chats need Conversation memoryNotes the model keeps on every token so far, so it doesn’t have to reread the conversation. It grows as the chat gets longer.Read the full entry, and on a small machine that runs out before the model’s limit.
- Follow you to your phone. Unless you set up a home server, your model stays on the computer it runs on.
Using both
Most people who run models at home also use the cloud. A common split:
- Local for private documents, routine rewriting and summarising, sorting files, and anything you do hundreds of times a day.
- Cloud for the hard problems, long projects, and questions about what is happening now.
Many apps can use either: Ollama runs a Local APIA way for other apps on your computer, or your network, to send requests to your model.Read the full entry that coding, note and chat apps connect to, so switching costs you nothing.
See what your computer can run · Run your first model in about 30 minutes