Apple’s first 2nm chip is now on sale. The M6 14-inch MacBook Pro and M6 iMac went on sale on 6 November 2026, after the Mac mini got the same chip in September.
Apple says the M6 is up to 25% faster than the M5 for on-device AI. The question I care about is simpler: which local models can you actually run on it, and is it worth buying for AI?
- Key specs: 12-core CPU, 12-core GPU with Neural Accelerators, dual 16-core Neural Engine, up to 32 GB unified memory, up to 170 GB/s bandwidth.
- Prices (US): MacBook Pro M6 from $1,999 (16 GB, 1 TB). iMac M6 from $1,499. Mac mini M6 from $899.
- 16 GB: 8B-12B models comfortably (Gemma 4 12B, MiMo-V2.6 9B). Tight for anything bigger.
- 32 GB: 27B-31B models at Q4 (Qwen3.8 27B, Gemma 4 31B) and 35-40B MoE models.
- Speed: 170 GB/s caps an 18 GB model at roughly 9 tokens/s (estimate). MoE models with about 3B active feel much faster.
- My advice: buy the 32 GB MacBook Pro or Mac mini for local AI. For 70B-class models you need an M5 Max or M5 Ultra.
M6 Specs That Matter for AI
For local AI, two numbers matter most: memory capacity decides which models fit, and memory bandwidth decides how fast they generate text.
| Spec | M6 | Why it matters for AI |
|---|---|---|
| Process | TSMC 2nm | Better performance per watt, longer battery under load |
| CPU | 12 cores (2 super, 4 performance, 6 efficiency) | Helps CPU-side work like tokenising and app overhead |
| GPU | 12 cores, a Neural Accelerator in each | Runs LLMs in llama.cpp, MLX and Ollama; faster prompt processing |
| Neural Engine | Dual 16-core, up to 1.9x M5 | Used by Apple Intelligence and Core ML apps, rarely by llama.cpp |
| Unified memory | 16 GB, 24 GB or 32 GB | Decides the biggest model you can load |
| Bandwidth | 153 GB/s (16 GB) / 170 GB/s (24-32 GB) | Decides tokens per second |
The bandwidth split comes from The Register: the 16 GB model runs at 153 GB/s, and the 24 GB and 32 GB options get the full 170 GB/s.
Note: Apple’s “up to 25% faster AI” and “1.3x faster time to first token” claims are about prompt processing, the part before the first word appears. Token generation speed is limited by bandwidth, which only rose about 11% over the M5.
M6 MacBook Pro, iMac and Mac mini Compared
Three Macs use the base M6. They have the same chip but different memory limits.
| Mac | CPU / GPU | Memory options | Starting price (US) | On sale |
|---|---|---|---|---|
| MacBook Pro 14-inch M6 | 12 / 12 | 16, 24 or 32 GB | $1,999 (16 GB, 1 TB) | 6 Nov 2026 |
| iMac M6, two ports | 10 / 10 | 16 or 24 GB | $1,499 (16 GB, 256 GB) | 6 Nov 2026 |
| iMac M6, four ports | 12 / 12 | 16, 24 or 32 GB | $1,699 (16 GB, 256 GB) | 6 Nov 2026 |
| Mac mini M6 | 12 / 12 | 16 to 32 GB | $899 (16 GB) | 22 Sep 2026 |
Warning: The $1,499 iMac has a cut-down 10-core GPU and tops out at 24 GB. For local AI, skip it and get the four-port model with 32 GB.
Other MacBook Pro M6 details from Apple: up to 24 hours of battery, Wi-Fi 7, Bluetooth 6, three Thunderbolt 4 ports and HDMI. Apple says performance is the same on battery and plugged in.
M6 vs M5 vs M5 Pro for AI
If you’re choosing between generations, here’s the comparison that matters for running models.
| M5 | M6 | M5 Pro | |
|---|---|---|---|
| CPU / GPU cores | 10 / 10 | 12 / 12 | up to 18 / 20 |
| Max memory | 32 GB | 32 GB | 64 GB |
| Bandwidth | 153 GB/s | up to 170 GB/s | 307 GB/s |
| Neural Engine | 16-core | Dual 16-core | 16-core |
| Best model size | up to ~30B Q4 | up to ~30B Q4 | up to ~70B Q4 (tight) |
| Cheapest way to get it | Discounted MacBook Pro | Mac mini $899 | Mac mini $1,699 |
Apple’s M6 Pro and M6 Max aren’t out. Macworld says they’re rumoured for 2027.
Which LLMs Fit in 16, 24 and 32 GB
macOS doesn’t give the GPU all of your unified memory. By default it keeps roughly a quarter to a third for the system. So plan on about 10-11 GB usable on 16 GB, 16-17 GB on 24 GB and 21-24 GB on 32 GB.
Download sizes below are from each model’s Ollama page. Add 1-2 GB for a normal context.
16 GB M6
- Gemma 4 12B (about 8 GB): the best all-rounder, with vision and a 256K context window.
- MiMo-V2.6 9B (6.6-8.7 GB): strong for coding and agents.
- Granite 4.2 8B (5.3 GB) and Bonsai 2 27B (8.3 GB, ternary).
- Skip: Qwen3.8 27B (18 GB) won’t fit.
24 GB M6
- Everything from the 16 GB list with long contexts.
- Gemma 4 26B MoE (16-19 GB, about 3.8B active) at a smaller quant.
- Qwen3.8 27B only with a smaller quant (Q3) and a short context.
32 GB M6
- Qwen3.8 27B (18 GB): my first pick, with a 16K-32K context.
- Gemma 4 31B (19-20 GB): best for maths and reasoning.
- Muse Glimmer 30B (18 GB) for agents, and Laguna XS 2.1 (20 GB).
- Flint 40B (23 GB, about 2.5B active): fits after raising the GPU limit, and runs fast because it’s MoE.
For the full list by memory size, see my best Ollama models guide and how much VRAM you need for AI.
How Fast Will It Be?
A simple rule works well: tokens per second is at most bandwidth divided by model size. Real numbers are usually 60-80% of that ceiling.
| Model (download size) | M6 ceiling (170 GB/s) | M5 Pro ceiling (307 GB/s) | RTX 5080 ceiling (960 GB/s) |
|---|---|---|---|
| Gemma 4 12B (8 GB) | ~21 tok/s | ~38 tok/s | ~120 tok/s |
| Qwen3.8 27B (18 GB) | ~9 tok/s | ~17 tok/s | won’t fit 16 GB fully |
| Gemma 4 31B (20 GB) | ~8 tok/s | ~15 tok/s | won’t fit 16 GB fully |
| Flint 40B MoE (~2.5B active) | much faster than dense, often 30+ tok/s | faster still | needs offload |
Note: These are ceilings I worked out from bandwidth and file size, not benchmarks I ran. I’ll add measured numbers once reliable M6 tests are published.
Tip: On a base M6, MoE models are the best way to get big-model quality at a usable speed. Only the active experts are read for each token.
Image and Video Generation on M6
Image and video models are compute-heavy, so the 12-core GPU is the limit here, not memory. The M6 can run ComfyUI and Draw Things, but expect it to be several times slower than a desktop RTX card.
- Image models: use MLX or Core ML builds where they exist (Draw Things), or GGUF and FP8 files in ComfyUI.
- Video models: possible on 32 GB with small GGUF quants, but slow. I’d rent a cloud GPU instead.
- CUDA-only nodes in ComfyUI won’t work on any Mac.
For heavy image and video work, compare with my cheapest cloud GPU guide.
Set Up Local AI on an M6 Mac
The fastest way to get going is Ollama or LM Studio. Both use the GPU on Apple silicon automatically.
- Install Ollama. Download it from ollama.com, open the app, and it adds the
ollamacommand. Full steps are in my Ollama install guide. - Pull a model that fits. On 16 GB run
ollama pull gemma4:12b. On 32 GB tryollama pull qwen3.8:27b. - Chat. Run
ollama run gemma4:12b, or use the Ollama app window. - Try MLX for speed. LM Studio can download MLX versions of many models, which are often faster on Macs than GGUF. Qwen3.8 also has a
27b-mlxtag in Ollama. - Add a nicer interface if you want one, such as Open WebUI.
Raise the GPU Memory Limit
If a model is just too big, you can let the GPU use more unified memory. This is a common trick on Apple silicon and resets when you restart.
- Open Terminal.
- Set the limit in MB. On a 32 GB Mac,
sudo sysctl iogpu.wired_limit_mb=26624allows about 26 GB. - Restart Ollama or LM Studio so it sees the new limit.
Warning: Leave at least 4-6 GB for macOS. If you push the limit too high, the Mac swaps heavily or freezes. Close browsers and other apps before loading a big model.
Which M6 Mac Should You Buy?
| You want… | Buy | Why |
|---|---|---|
| Cheapest Mac for local AI | Mac mini M6, 32 GB | Same chip and memory as the MacBook Pro for much less |
| A laptop that runs 27B-31B models | MacBook Pro M6, 32 GB | Full bandwidth, 24-hour battery, quiet |
| An all-in-one desktop | iMac M6 four-port, 32 GB | The two-port iMac stops at 24 GB |
| 70B models or long contexts | Mac mini M5 Pro 64 GB or Mac Studio M5 Max | More memory and roughly 2-3.5x the bandwidth |
| Image/video generation, training | A PC with an NVIDIA GPU | CUDA support and far more compute |
For the bigger picture, read Mac vs PC for local AI and my best laptop for local AI picks.
Troubleshooting
The model is very slow
- Check memory pressure in Activity Monitor. Yellow or red means the model doesn’t fit and macOS is swapping.
- Use a smaller quant (Q4_K_M instead of Q6) or a smaller model.
- Lower the context to 8K-16K. A huge context eats memory.
The model won’t load or Ollama falls back to CPU
- Model too big for the GPU limit: run
ollama psto see the GPU/CPU split. Raise the limit as shown above, or pick a smaller file. - Old version: update Ollama or LM Studio. New models often need the latest release.
The fans spin up or the battery drains fast
Long generations keep the GPU at full load. Plug in for long runs, and use the High Power mode setting only when you need it.
Apple Intelligence uses memory too
Apple’s own on-device models take some memory when active. On a 16 GB Mac, turning off features you don’t use can free a little room.
Frequently Asked Questions
Is the M6 MacBook Pro good for local LLMs?
Yes for 8B-31B models if you get 32 GB. It’s not the right choice for 70B models, because the base M6 tops out at 32 GB and 170 GB/s.
How much RAM do I need on an M6 Mac for AI?
Get 32 GB. 16 GB limits you to models around 12B, and 24 GB is awkward for 27B-31B models.
Is the M6 faster than the M5 for AI?
Apple says up to 25% faster for on-device AI and 1.3x faster time to first token. Token generation should improve less, roughly in line with the bandwidth rise from 153 GB/s to 170 GB/s.
Should I buy the M6 Mac mini or the M6 MacBook Pro for AI?
For a desk setup, the Mac mini M6 with 32 GB gives you the same AI performance for much less money. Buy the MacBook Pro only if you need a laptop.
Can the M6 run Stable Diffusion or ComfyUI?
Yes, with MLX, Core ML apps like Draw Things, or ComfyUI with GGUF/FP8 models. It will be much slower than an NVIDIA desktop GPU, and CUDA-only nodes won’t work.
Is there an M6 Pro or M6 Max?
Not yet. Only the base M6 is out, in the MacBook Pro, iMac and Mac mini. Reports point to M6 Pro and Max chips in 2027.
Which iMac M6 is better for AI?
The four-port model. It has the full 12-core GPU and goes up to 32 GB, while the $1,499 two-port model has a 10-core GPU and a 24 GB limit.
Final Thoughts
The M6 is a nice step up, especially for battery life and prompt processing. But for local AI the deciding numbers barely moved: 32 GB maximum and 170 GB/s.
If you buy one for AI, get 32 GB and lean on MoE models. If you want 70B-class models, look at the M5 Pro or M5 Max instead, or build a PC with the best GPU for local LLMs.