LocalLabWhat runs on my Mac?
Memory decides which AI models fit on a Mac; the chip decides how fast they run. Here is every model LocalLab AI offers, graded for seven typical Macs by the same Smart Fit the app uses.
By memory
8 GB — MacBook Air, M1
- Chat runs well up to Qwen3.5 2B; Qwen3.5 4B also runs, more slowly or with little room to spare.
- Runs well: stem separation (HTDemucs v4), text recognition (GLM-OCR).
- Fits, with the time set by the chip: photo upscaling (VOSR2).
- Runs, but slowly: text to speech (Qwen3-TTS 0.6B), transcription (Whisper large-v3 turbo).
- Needs more memory: image generation (FLUX.1 schnell), video generation (FastWan 2.2), music generation (MiniMax Music 3), text to speech (Qwen3-TTS 1.7B), video upscaling (SeedVR2 3B).
16 GB — MacBook Air, M2
- Chat runs well up to Qwen3.5 4B; Qwen3.5 9B also runs, more slowly or with little room to spare.
- Runs well: text to speech (Qwen3-TTS 0.6B), text to speech (Qwen3-TTS 1.7B), stem separation (HTDemucs v4), transcription (Whisper large-v3 turbo), text recognition (GLM-OCR).
- Fits, with the time set by the chip: video generation (FastWan 2.2), video upscaling (SeedVR2 3B), photo upscaling (VOSR2).
- Runs, but slowly: image generation (FLUX.1 schnell), music generation (MiniMax Music 3).
24 GB — MacBook Air, M4
- Chat runs well up to Qwen3.5 4B; Qwen3.5 9B, Qwen3.5 27B also run, more slowly or with little room to spare.
- Runs well: music generation (MiniMax Music 3), text to speech (Qwen3-TTS 0.6B), text to speech (Qwen3-TTS 1.7B), stem separation (HTDemucs v4), transcription (Whisper large-v3 turbo), text recognition (GLM-OCR).
- Fits, with the time set by the chip: video generation (FastWan 2.2), video upscaling (SeedVR2 3B), photo upscaling (VOSR2).
- Runs, but slowly: image generation (FLUX.1 schnell).
32 GB — MacBook Pro, M2 Max
- Chat runs well up to Qwen3.5 9B; Qwen3.5 27B, Qwen3.5 35B-A3B also run, more slowly or with little room to spare.
- Runs well: image generation (FLUX.1 schnell), music generation (MiniMax Music 3), text to speech (Qwen3-TTS 0.6B), text to speech (Qwen3-TTS 1.7B), stem separation (HTDemucs v4), transcription (Whisper large-v3 turbo), text recognition (GLM-OCR).
- Fits, with the time set by the chip: video generation (FastWan 2.2), video upscaling (SeedVR2 3B), photo upscaling (VOSR2).
48 GB — MacBook Pro, M4 Pro
- Chat runs well up to Qwen3.5 35B-A3B; Qwen3.5 27B also runs, more slowly or with little room to spare.
- Runs well: image generation (FLUX.1 schnell), music generation (MiniMax Music 3), text to speech (Qwen3-TTS 0.6B), text to speech (Qwen3-TTS 1.7B), stem separation (HTDemucs v4), transcription (Whisper large-v3 turbo), text recognition (GLM-OCR).
- Fits, with the time set by the chip: video generation (FastWan 2.2), video upscaling (SeedVR2 3B), photo upscaling (VOSR2).
64 GB — Mac Studio, M4 Max
- Chat runs well up to Qwen3.5 35B-A3B.
- Runs well: image generation (FLUX.1 schnell), music generation (MiniMax Music 3), text to speech (Qwen3-TTS 0.6B), text to speech (Qwen3-TTS 1.7B), stem separation (HTDemucs v4), transcription (Whisper large-v3 turbo), text recognition (GLM-OCR).
- Fits, with the time set by the chip: video generation (FastWan 2.2), video upscaling (SeedVR2 3B), photo upscaling (VOSR2).
128 GB — Mac Studio, M4 Max
- Chat runs well up to Qwen3.5 122B-A10B.
- Runs well: image generation (FLUX.1 schnell), music generation (MiniMax Music 3), text to speech (Qwen3-TTS 0.6B), text to speech (Qwen3-TTS 1.7B), stem separation (HTDemucs v4), transcription (Whisper large-v3 turbo), text recognition (GLM-OCR).
- Fits, with the time set by the chip: video generation (FastWan 2.2), video upscaling (SeedVR2 3B), photo upscaling (VOSR2).
Every model, every Mac
- Runs well
- Fits — time depends on the chip
- Runs, slowly
- Tight — close other apps
- Not enough memory
| Tool · model | 8 GBMacBook Air, M1 | 16 GBMacBook Air, M2 | 24 GBMacBook Air, M4 | 32 GBMacBook Pro, M2 Max | 48 GBMacBook Pro, M4 Pro | 64 GBMac Studio, M4 Max | 128 GBMac Studio, M4 Max |
|---|---|---|---|---|---|---|---|
| Chat · Qwen3.5 0.8B0.6 GB download | Runs well: ~59 tok/s · 32k | Runs well: ~84 tok/s · 32k | Runs well: ~98 tok/s · 32k | Runs well: ~244 tok/s · 32k | Runs well: ~188 tok/s · 32k | Runs well: ~293 tok/s · 32k | Runs well: ~293 tok/s · 32k |
| Chat · Qwen3.5 2B1.7 GB download | Runs well: ~23 tok/s · 32k | Runs well: ~33 tok/s · 32k | Runs well: ~39 tok/s · 32k | Runs well: ~115 tok/s · 32k | Runs well: ~83 tok/s · 32k | Runs well: ~148 tok/s · 32k | Runs well: ~148 tok/s · 32k |
| Chat · Qwen3.5 4B3 GB download | Runs, slowly: ~13 tok/s · 4k | Runs well: ~19 tok/s · 32k | Runs well: ~23 tok/s · 32k | Runs well: ~71 tok/s · 32k | Runs well: ~50 tok/s · 32k | Runs well: ~93 tok/s · 32k | Runs well: ~93 tok/s · 32k |
| Chat · Qwen3.5 9B6 GB download | Not enough memory: needs 6.7 GB | Runs, slowly: ~10 tok/s · 32k | Runs, slowly: ~12 tok/s · 32k | Runs well: ~38 tok/s · 32k | Runs well: ~26 tok/s · 32k | Runs well: ~51 tok/s · 32k | Runs well: ~51 tok/s · 32k |
| Chat · Qwen3.5 27B16.1 GB download | Not enough memory: needs 16.2 GB | Not enough memory: needs 16.2 GB | Runs, slowly: ~4 tok/s · 4k | Runs, slowly: ~15 tok/s · 8k | Runs, slowly: ~10 tok/s · 32k | Runs well: ~20 tok/s · 32k | Runs well: ~20 tok/s · 32k |
| Chat · Qwen3.5 35B-A3B20.4 GB download | Not enough memory: needs 20.1 GB | Not enough memory: needs 20.1 GB | Not enough memory: needs 20.1 GB | Tight — close other apps: ~68 tok/s · 4k | Runs well: ~48 tok/s · 32k | Runs well: ~89 tok/s · 32k | Runs well: ~89 tok/s · 32k |
| Chat · Qwen3.5 122B-A10B69.6 GB download | Not enough memory: needs 65.9 GB | Not enough memory: needs 65.9 GB | Not enough memory: needs 65.9 GB | Not enough memory: needs 65.9 GB | Not enough memory: needs 65.9 GB | Not enough memory: needs 65.9 GB | Runs well: ~30 tok/s · 32k |
| Image · FLUX.1 schnell9.9 GB download | Not enough memory: needs 7.3 GB | Runs, slowly: about 4 min an image | Runs, slowly: about 3 min an image | Runs well: about 75 s an image | Runs well: about 90 s an image | Runs well: about 45 s an image | Runs well: about 45 s an image |
| Video · FastWan 2.218.9 GB download | Not enough memory: needs 10.3 GB | Fits — time depends on the chip: | Fits — time depends on the chip: | Fits — time depends on the chip: | Fits — time depends on the chip: | Fits — time depends on the chip: | Fits — time depends on the chip: |
| Music · MiniMax Music 39.2 GB download | Not enough memory: needs 6.9 GB | Runs, slowly: about 20 min for 30 s | Runs well: about 11 min for 30 s | Runs well: about 5 min for 30 s | Runs well: about 5 min for 30 s | Runs well: about 3 min for 30 s | Runs well: about 3 min for 30 s |
| Voice · Qwen3-TTS 0.6B2.5 GB download | Runs, slowly: about 3 min for a page | Runs well: about 2 min for a page | Runs well: about 2 min for a page | Runs well: about 45 s for a page | Runs well: about 54 s for a page | Runs well: about 27 s for a page | Runs well: about 27 s for a page |
| Voice · Qwen3-TTS 1.7B4.5 GB download | Not enough memory: needs 5.6 GB | Runs well: about 3 min for a page | Runs well: about 2 min for a page | Runs well: about 53 s for a page | Runs well: about 63 s for a page | Runs well: about 31 s for a page | Runs well: about 31 s for a page |
| Upscale · SeedVR2 3B (videos)4.7 GB download | Not enough memory: needs 11.1 GB | Fits — time depends on the chip: 5 s of 384p video to 768p | Fits — time depends on the chip: 5 s of 384p video to 768p | Fits — time depends on the chip: 5 s of 384p video to 768p | Fits — time depends on the chip: 5 s of 384p video to 768p | Fits — time depends on the chip: 5 s of 384p video to 768p | Fits — time depends on the chip: 5 s of 384p video to 768p |
| Upscale · VOSR2 (photos)3.3 GB download | Fits — time depends on the chip: a 1024×1024 image to 2048×2048 | Fits — time depends on the chip: a 1024×1024 image to 2048×2048 | Fits — time depends on the chip: a 1024×1024 image to 2048×2048 | Fits — time depends on the chip: a 1024×1024 image to 2048×2048 | Fits — time depends on the chip: a 1024×1024 image to 2048×2048 | Fits — time depends on the chip: a 1024×1024 image to 2048×2048 | Fits — time depends on the chip: a 1024×1024 image to 2048×2048 |
| Separate · HTDemucs v40.2 GB download | Runs well: about 33 s for 4 min | Runs well: about 24 s for 4 min | Runs well: about 19 s for 4 min | Runs well: about 8 s for 4 min | Runs well: about 9 s for 4 min | Runs well: about 5 s for 4 min | Runs well: about 5 s for 4 min |
| Transcribe · Whisper large-v3 turbo1.6 GB download | Runs, slowly: about 4 min for 30 min | Runs well: about 3 min for 30 min | Runs well: about 2 min for 30 min | Runs well: about 54 s for 30 min | Runs well: about 63 s for 30 min | Runs well: about 32 s for 30 min | Runs well: about 32 s for 30 min |
| Read text · GLM-OCR2.7 GB download | Runs well: a few seconds a page | Runs well: a few seconds a page | Runs well: a few seconds a page | Runs well: a few seconds a page | Runs well: a few seconds a page | Runs well: a few seconds a page | Runs well: a few seconds a page |
Other apps closed: each Mac's memory less 2.5 GB for macOS. Speeds are the owner's M2 Max measurements carried to each chip by its specifications. Chat is graded at the longest context that fits well; the other tools on a typical job: a page of text, a 30-second song, half an hour of speech, a 4-minute track, a 1024 × 1024 picture, a 5-second clip. Graded 2026-10-19.
How LocalLab AI decides
On first launch, LocalLab AI runs a few seconds of speed tests on your Mac. Then, before anything downloads, it grades every model and job for your Mac — the memory it needs against the memory free right now, and how long it should take — and shows the arithmetic. Chat picks the best model your Mac runs well; long jobs wait in a queue until there is room.
Questions
Can an 8 GB Mac run AI models locally?
Yes. On an 8 GB Mac with Apple silicon, LocalLab AI runs chat with the smaller Qwen3.5 models, text to speech, transcription with Whisper, stem separation, text recognition and photo upscaling. Image, video and music generation need 16 GB or more.
How much memory do I need to run a 27B language model on a Mac?
Qwen3.5 27B needs about 16 GB for itself, so it runs on Macs with 24 GB or more — slowly on 24 GB, and comfortably from 64 GB. On a 32 GB Mac, the 35B-A3B mixture-of-experts model is much faster, because it reads only 3B of its parameters for each token.
What matters more for local AI on a Mac: memory or the chip?
Memory decides what fits; the chip decides how fast it runs. Language models generate text about as fast as the chip's memory bandwidth allows, so a Pro, Max or Ultra chip writes several times faster than a base chip with the same memory.
Does LocalLab AI run on Intel Macs?
No. It needs a Mac with Apple silicon (M1 or later) and macOS 14 or later, because its models run on the Apple GPU with Apple's MLX.
Why does my Mac get a different grade from this page?
This page assumes other apps are closed. LocalLab AI grades every model against the memory free on your Mac right now and measures your Mac's speed on first launch, so a busy Mac may see more yellow, and a measured one more accurate times.