How much RAM a Mac needs for Whisper and a local language model depends on whether both have to be loaded at once. A Mac with 8 GB can run a small speech model or a small language model, one at a time, with some swapping. 16 GB is the realistic minimum for Whisper large-v3-turbo plus a 7B or 8B model at 4-bit, as long as you load them in turn; 24 to 32 GB keeps both loaded with the machine still pleasant to use.
Below: what each model occupies, why long transcripts cost extra, and how to read your Mac's memory. If you would rather not size models at all, the last section shows what ThinkScribe for Mac needs instead.
Why unified memory changes the question
On an Intel Mac with a discrete graphics card, the card had its own memory and the model had to fit there. Apple silicon uses unified memory: the CPU, GPU and Neural Engine work over one shared pool, and nothing is copied between them. A model loaded once is visible to every processor.
That is also the catch. The model competes with Safari, your chat app, your editor and macOS itself for the same gigabytes, and there is no separate graphics memory figure to check. When a speech model and a language model are both loaded on a 16 GB Mac, the rest of the system gets what is left, and macOS starts compressing and swapping memory well before the number reaches zero.
What Whisper occupies
Whisper comes in several sizes, published as files at full 16-bit precision and in compressed (quantized) versions. The "loaded" column is an approximate figure for the full-precision model while it runs. Compressed files load into proportionally less, but the working memory (audio buffers, encoder activations, the decoder's cache) does not shrink with compression, so treat everything as approximate.
| Model | 16-bit file | Compressed file | Loaded (approx.) | Parameters |
|---|---|---|---|---|
| tiny | 78 MB | about 32 MB | about 270 MB | 39 M |
| base | 148 MB | about 60 MB | about 390 MB | 74 M |
| small | 488 MB | about 190 MB | about 850 MB | 244 M |
| medium | 1.53 GB | about 540 MB | about 2.1 GB | 769 M |
| large-v3 | 3.1 GB | about 1.1 GB | about 3.9 GB | 1,550 M |
| large-v3-turbo | 1.62 GB | 574 to 874 MB | about 2 to 2.5 GB (estimate) | 809 M |
For the turbo model the loaded figure is an estimate, following the other rows, where memory in use runs roughly 0.5 to 1 GB above the file size. Other open speech models have their own profiles; Whisper vs Parakeet and the open speech recognition models of 2026 compare them.
What a local language model occupies
Language models scale with parameter count and bits per weight. The rule of thumb is about 0.55 to 0.65 bytes per parameter at 4-bit (the extra fraction is scales, embeddings and layers kept at higher precision), just over 1 byte per parameter at 8-bit, and 2 bytes at 16-bit.
| Parameters | 4-bit (approx.) | 8-bit (approx.) | 16-bit |
|---|---|---|---|
| 0.5 B | 400 MB | 530 MB | 1.0 GB |
| 1.5 to 2 B | 1.0 GB | 1.6 GB | 3.1 GB |
| 3 to 4 B | 1.9 to 3.3 GB | 3.3 to 5.0 GB | 6.2 to 8.6 GB |
| 7 to 8 B | 4.7 GB | 8.1 GB | 15 GB |
| 12 to 14 B | 8.1 to 9.0 GB | 13 to 16 GB | 24 to 30 GB |
A common guideline is at least 8 GB of available memory for 7B models, 16 GB for 13B models and 32 GB for 33B models. Note the word available. A 4.7 GB model on an 8 GB Mac works because macOS compresses and swaps everything else; it is not comfortable.
The KV cache: why a long transcript costs more than you expect
The weights are the fixed cost. The variable cost is the key-value (KV) cache, which stores attention state for every token in the context and grows in step with prompt length. Its size per token is: 2 (keys and values) × layers × KV heads × head dimension × bytes per value.
Take a typical 7B model with 28 layers, 4 key-value heads and 128-dimensional heads. At 16-bit that is 2 × 28 × 4 × 128 × 2 = 57,344 bytes, about 56 KB per token. A one-hour meeting transcript is roughly 9,000 words, around 12,000 tokens, so the cache for that prompt alone is about 670 MB before the model writes a word. Models with more KV heads pay more: 32 layers with 8 KV heads needs about 128 KB per token, around 1.5 GB for the same transcript and 4 GB at a 32,000-token context, close to the weights themselves.
This is why "the 4-bit model is 4.7 GB" is the wrong number to plan with for transcript work. Plan with weights plus cache at your longest realistic prompt, plus a few hundred megabytes for the runtime.
Running both at once
A transcribe-then-summarize workflow has two peaks. If the speech model is unloaded before the language model loads, your budget is the larger of the two. If both stay loaded, it is the sum: speech model, plus language model weights, plus KV cache at the transcript's length. On a 16 GB Mac, large-v3-turbo (about 2.5 GB) plus a 7B 4-bit model (4.7 GB) plus a one-hour cache (0.7 to 1.5 GB) is 8 to 9 GB for the models alone, leaving about 7 GB for macOS and everything else. It runs; it is not roomy.
The first run is slower and hungrier
Models that run on the Neural Engine are compiled for your specific chip the first time they load. Apple's Core ML Tools documentation describes this as compilation "for backend device specialization, such as for the Neural Engine (NE), which may take a few seconds or even minutes for large models", with an example of 14.7 seconds for a first load against 0.1 seconds for the second.
Two consequences for memory. The compile is a temporary spike on top of the model's normal footprint, so a Mac that is fine on the second run can show yellow memory pressure on the first. And a macOS update can clear the cached result and trigger the compile again.
Practical budgets by Mac
8 GB. One model at a time: Whisper small, a compressed medium, or compressed large-v3-turbo as the ceiling, and a 1B to 3B language model at 4-bit. Prefer compressed files, load one model at a time, split long transcripts so no prompt exceeds a few thousand tokens, and quit the browser during long jobs.
16 GB. The comfortable single-model tier. large-v3-turbo runs with room to spare, and a 7B or 8B model at 4-bit with an 8,000 to 16,000-token context is fine. Both loaded at once works for one-hour transcripts but leaves little headroom; loading in turn is better.
24 to 32 GB. Both loaded at once, long contexts, and you can keep working. A 12B to 14B model at 4-bit alongside large-v3 is realistic at 32 GB.
64 GB and above. 30B-class models at 4-bit and several speech models loaded together. Memory stops being the constraint; time per token does.
How to read your Mac's memory
Activity Monitor's Memory tab is the honest first look. Apple's guidance for the memory pressure graph: green means your Mac is using its memory efficiently, yellow means it might eventually need more, red means it needs more. Watch Swap Used and Compressed rather than Memory Used, which is nearly always high on a Mac.
When memory runs short, macOS compresses inactive memory, then swaps to the SSD. Long before any warning appears, a language model slows to a crawl because its weights are read back from disk on every step.
How to transcribe and summarize on a Mac without sizing models
If your goal is the transcript and the summary rather than the tinkering, ThinkScribe for Mac does both with no model sizes to choose.
- Check your Mac: ThinkScribe for Mac needs Apple silicon and macOS 26 or later (Apple menu > About This Mac shows both).
- Install it and record (⌘R) or import a file (⌘I). ThinkScribe sets up the speech model for you, and the transcript is made on your Mac when you stop.
- Choose the AI engine in Settings > AI Assistant: Apple Intelligence (provided by Apple, on Macs that support it) or ThinkScribe AI (a one-time download that runs on your device).
- Summarize it, or ask for Meeting Minutes, an Executive Brief or a Mind Map, or ask questions about the recording.
On iPhone, memory does shape a few features. Mind Map needs a recent iPhone with more memory and is hidden on models with less. Live speaker labels while recording are available on iPhones with 8 GB of memory, off by default. Transcription, speaker labels after you stop, summaries and Chat work on the supported iPhones. The complete guide to voice to text covers every device.
Frequently asked questions
How much RAM does Whisper need on a Mac?
Roughly 270 MB for tiny, 850 MB for small, 2.1 GB for medium and 3.9 GB for large-v3 at full precision, with compressed files loading into proportionally less. On 8 GB use small, a compressed medium or compressed large-v3-turbo; on 16 GB any size runs comfortably as the only large model loaded.
Can an 8 GB Mac run a local LLM and Whisper?
One at a time. A 1B to 3B model at 4-bit plus a compressed small or medium Whisper is workable if you unload each before loading the next and keep prompts short. A 7B model on 8 GB means heavy swapping, and alongside a speech model it is not practical.
Does the KV cache really matter for summarizing a transcript?
Yes, because the whole transcript goes into the prompt. At 16-bit the cache takes roughly 56 to 128 KB per token depending on the model's design, so a one-hour meeting of around 12,000 tokens adds 0.7 to 1.5 GB on top of the weights.
What Mac do I need for ThinkScribe?
A Mac with Apple silicon and macOS 26 or later. There are no model sizes to pick: ThinkScribe sets up the speech model for you, and you choose Apple Intelligence or ThinkScribe AI for summaries and Chat.
