Mac Mini M6: Local AI

Testing Apple's AI-Focused Mac

Apple is pushing their 'Apple Intelligence' particularly hard, perhaps out of a need for Siri to be well... a bit faster on-device.

Much like their new mobile A20 Pro, the M6 has two 16-core Neural Engines. That hardware only helps software written for it, and LM Studio on Apple Silicon typically runs through Metal or MLX, so it's worth being clear this is not a Neural Engine benchmark.

What good looks like

I can provide numbers, but for those of you looking for what they mean, this is my subjective view. Realistically, many people would set the lines somewhere else here. Token generation is how fast the response trickles in after encoding your message and any other context.

Anything below 10 tok/s is too slow for chat, for better or worse, skim-reading an AI response first to get a feel for how valuable it is has become second-nature and anything slower than 10 tok/s feels lethargic.

For agentic or tool-using work where I'm leaning on reasoning and planning, I want at least 30 tok/s. Under that, the model is thinking in a loop I can feel a bit too much.

In an ideal world I'd aim for above 50 tok/s, this is closer to the kind of speed you might expect using models like Opus/Fable outside of the fast lane.

Time to first token ultimately still matters, and so does whether the model even fits. The gold standard at the time of writing, subjectively of course would be Qwen3.8-27B, so getting acceptable performance on the M6 with that would be a win.

How I tested

The same prompt was used on every machine that can load the file:

If Apple were to release a budget laptop leveraging the chip of the Pro, do you think it would be a functional laptop with reasonable performance for every day tasks? What specification do you anticipate the laptop might have?

LFM 2.5 1.2B and Qwen 4B Thinking (2507) are 8-bit. Gemma 4 12B and Qwen 3.8 27B are 4-bit. This is because an 8-bit 27B is (you'll never guess) around the region of 30GB. It's not going to fit on the 24GB M6, or most of the other machines here. Even still, in my experience a 4-bit quant of Qwen3.8-27B is plenty capable.

On 8-bit Qwen 4B, the Neo is under my chat floor at 8.95 tok/s. The base M4 Mini is chat-usable at 20.7, and the binned M3 Pro lands at 27.8, which is fine for reading and just shy of how fast I want a tool loop. The M6 Mini clears my agentic line at 31.47 tok/s with a first token in 0.64s, which is a 52% gain over the M4 and enough to make the small thinking model genuinely usable in a loop. The M1 Max is still ahead at 51.6, which is what 400GB/s of first-generation Max bandwidth buys. The 5070 is in another class at 85.8, with a first token in 0.03s, but it is a different runtime and it runs out of VRAM long before a 64GB Mac does.

LFM 2.5 is the easy model. The Neo holds 37.7 tok/s, the M4 Mini 71.2, the M6 Mini 111.8, the M3 Pro 98. Gemma 4 12B is the 4-bit QAT column: the M4 Mini sits on my chat floor at 10.7, the M3 Pro is more comfortable at 16.9, the M6 lands at 19.0, the M1 Max holds 22.9, and the 5070 is agentic-comfortable at 54.2. I would not line that up against the 8-bit 4B numbers and call it a clean size scaling.

The pattern across the chart is the one you would expect from a bandwidth-bound workload. The M6 beats every other Apple machine here on the small models, where compute and first-token latency dominate, and loses to the M1 Max on the larger ones, where 170GB/s runs into 400GB/s. A generational CPU leap does not buy you memory bandwidth.

GPU power during these runs stayed modest: about 5.4W average on Qwen 4B, 7.7W on Gemma 12B with a 22W peak, and 5.7W on the 27B. This is a machine you can leave holding a model.

Larger models

Qwen 3.8 27B 4-bit QAT MLX is the capacity check. It only belongs on the 24GB M6 and the 64GB M1 Max. Everyone else is out. Medium reasoning, same prompt.

The 27B does load on 24GB, with little room to spare. It's also the clearest limit on this machine, as 9.42 tok/s puts it just under my chat floor and the 12.28 second wait for a first token is excruciating given how short the prompt was. The M1 Max produces tokens only a little faster at 11.9, but starts answering in 1.81s, because it is not living right at the edge of its memory.

So the honest answer on the 24GB M6 and a 27B model is, it runs, and you will not want to use it. The models this machine is genuinely good at are the ones up to about 12B, where it clears my floors comfortably and sips power doing it. That's still an impressive result, and I would wager that the 32GB M6 Mac Mini actually performs well here.

My unit is 24GB LPDDR5X-10667 (up to 170GB/s). The 16GB M4 Mini is on slower LPDDR5X-9600, so I'd imagine it performs perform closely to the M3 Pro.

Gaming BenchmarksCyberpunk 2077 at native 1080p, with the same Windows desktop as the discrete-GPU ceiling.