Mac mini M6: local AI test and review
The first 2 nm Mac — and the cheapest entry point to local AI from Apple. We look at what matters for running LLMs locally : memory, bandwidth, speed, price—and who it's (or isn't) the right purchase for.
Updated on 09/10/2026
Where to buy it
The button configuration is the one offered for purchase. On mini PCs, keep some memory available for the system and check the inference engine; macOS/MLX and CUDA are not interchangeable.
Affiliate links — QuelLLM may earn a commission on purchases, at no extra cost to you, which does not influence these recommendations (established independently). As an Amazon Associate, BestLLMfor earns from qualifying purchases.
Technical specifications
- Memory16, 24, or 32 GB unified memory (24 GB recommended for local AI)
- Bandwidth170 GB/s (+10% vs. M5)
- ComputeApple M6 — 1st 2 nm chip (TSMC): 12-core CPU (2 efficiency + 4 performance + 6 efficiency), 12-core GPU with Neural Accelerators (+30% GPU AI compute vs. M5), dual 16-core Neural Engine
- Power consumptionCompact, quiet desktop system
- Indicative price€1,049 (16 GB / 256 GB) · €1,489 (24 GB / 512 GB, the local AI configuration) · +€440 for 32 GB
- Largest model16 GB: 7-9B in Q4 · 24 GB: comfortable 14B + MoE 30B-A3B · 32 GB: 30B class is loadable but slow (~6-8 tok/s)—the real ceiling is 170 GB/s
Who is it for?
Advantages and limitations
✓ Strengths
- First Apple chip built on a 2 nm process — 12-core CPU, +30% GPU AI compute vs. M5
- Neural Accelerators in every GPU core: significantly accelerated prefill (prompt reading)
- The cheapest new Mac for trying out local AI (Ollama, LM Studio, MLX run perfectly)
- Lightweight MoE models (30B-A3B) are surprisingly usable: estimated ~40-60 tok/s on 24-32 GB
- Compact, quiet, energy-efficient — perfect as a primary machine that ALSO does AI
✗ Limitations
- 170 GB/s of bandwidth: half as much as an M5 Pro (307 GB/s) — dense models > 14B crawl
- 32 GB ceiling: neither 70B nor large MoE models—this is not the machine for large models
- Generational price increase: €1,049 in 16/256 (the Mac mini M4 started much lower)
What the M6 can run, tier by tier
At 170 GB/s, bandwidth—not memory—is the limiting factor for dense models. MoE (Mixture of Experts) models work around the issue: they read only their active parameters for each token. Order-of-magnitude speeds (Q4, MLX / llama.cpp).
| Configuration | What runs (Q4) | Estimated speed | Typical use |
|---|---|---|---|
| 16 GB—€1,049 | 7–9B (Qwen 3.5 9B, Gemma…) | ~25-30 tok/s | Chat, summary, lightweight copilot |
| 24 GB — €1,489 | 14B is comfortable; 30B-A3B MoE (~18 GB) | ~15-18 tok/s (14B) · ~40-60 tok/s (MoE) | Code copilot, personal RAG — the recommended configuration |
| 32 GB — €1,489 + €220 | Loadable 30B-class model (Qwen3-32B) | ~6-8 tok/s (30B dense) | Possible but frustrating—prefer MoE models, or an M5 Pro |
The real M6 bonus: the Neural Accelerators on each GPU core speed up prefilling — on a long prompt (RAG, large file), the wait before the first token drops noticeably compared with an M4.
M6 or M5 Pro: the €510 that changes everything
The same chassis is available in Mac mini M5 Pro at €1,999 (24 GB / 512 GB): 307 GB/s instead of 170, a 16-core GPU, and 48–64 GB tiers. In practice: a 14B model goes from ~15–18 to ~28–33 tok/s, the 30B class becomes genuinely usable, and 64 GB even makes a 70B model in Q4 possible (slow, but loadable). If your local AI budget exceeds €1,500, the question isn’t “M6 24 or 32 GB?” but “M6 24 GB or M5 Pro?” — and for AI, it’s the M5 Pro.
The memory upgrade from the M6 to 32 GB (+€440 total from €1,049) is the catalog trap: you pay for memory that the bandwidth cannot serve on dense models. That €440 funds half the gap to the M5 Pro.
Compared with the alternatives
| Machine | Memory | Bandwidth | Pricing | Using it properly |
|---|---|---|---|---|
| Mac mini M6 (24 GB) | 24 GB unified memory | 170 GB/s | 1 489 € | Primary macOS machine + secondary local AI |
| Mac mini M5 Pro | 24-64 GB unified | 307 GB/s | 1 999 € | Serious local AI on macOS (30B class) |
| PC + RTX 5060 Ti 16 GB | 16 GB VRAM | 448 GB/s | ≈ €800–€820 (card only) | Maximum speed per euro up to 14-24B (if you already have a PC) |
| Strix Halo Mini-PC 64 GB | 64 GB unified | 212 GB/s | ≈ 2 000 à 2 300 € | Maximum memory capacity under Windows/Linux (70B may run slowly) |
A dedicated 16 GB GPU remains unbeatable for raw speed on small models — but you need a PC around it, plus noise and ~200 W more at the wall.
Availability: what you need to know
- On sale since September 22, 2026 (preorders opened on August 25).
- Three configurations listed by French retailers: 16/256 (€1,049), 16/512, and 24/512 at €1,489 — the one you need for local AI.
- The 32 GB tier requires the Apple Store’s BTO (+€440 from the base configuration)—see our “M6 or M5 Pro” section before clicking.
- Same refresh on 08/25: the Mac mini M5 Pro (€1,999) and the Mac Studio M5 Max / Ultra (starting at €2,999).
Verdict
Frequently asked questions
Can the Mac mini M6 run a 70B LLM locally?
16 GB: 7–9B in Q4 · 24 GB: comfortable 14B + 30B-A3B MoE · 32 GB: 30B class can load but is slow (~6–8 tok/s)—the real ceiling is 170 GB/s.
How much does the Mac mini M6 cost?
Budget €1,049 (16 GB / 256 GB) · €1,489 (24 GB / 512 GB, the local AI configuration) · +€440 for 32 GB (from $899). Prices change quickly—check today’s price through the purchase links.
Who is the Mac mini M6 for?
Discover local AI on a budget with macOS: 7–14B models and lightweight MoE models, alongside everyday office and development use.
Is the Mac mini M6 enough for a local coding copilot?
Yes, at 24 GB: a 7–14B coding model (or a Qwen3-Coder-30B-A3B–type MoE) runs at comfortable speeds for autocompletion and chat in VS Code. That is exactly the scope of the Local Copilot kit. For heavy agentic workloads or the dense 30B class, move up to the M5 Pro.
Why doesn't the 2 nm chip make LLMs faster?
Text generation is limited by memory read speed (bandwidth), not computation: each token requires rereading all of the model's active parameters. The M6 mainly gains in prefill (+30% GPU AI compute, Neural Accelerators)—the wait before the first token—but its 170 GB/s caps the generation throughput of dense models.
16, 24, or 32 GB: which configuration should you choose?
24 GB (1,489 € for 24/512), without hesitation: the 16 GB tier limits you to 7–9B models, while the 32 GB tier costs 440 € for capacity that the bandwidth cannot keep up with. If you are considering 32 GB “for AI,” put that money toward the Mac mini M5 Pro instead.