M6 Mac Mini Benchmarks: Is It Worth Upgrading from the M4?
M6 vs M4 Mac mini benchmarks compared: CPU, GPU, SSD, memory bandwidth, and local AI performance to see if upgrading is worth it.
The short answer
Yes, if you run local AI models or compile code often. The M6 Mac mini beats the M4 by wide margins in prompt processing, token generation, browser performance, and Python execution, largely thanks to a new 12-core GPU with per-core neural accelerators, a second neural engine, and higher memory bandwidth. If you mostly browse, email, and stream video, the M4 still handles that fine and the M6 won’t feel transformative for those tasks.
TL;DR
- The M6 chip mixes three core types for the first time: two “super cores,” four performance cores, and six efficiency cores, replacing Apple’s old two-tier CPU design.
- Every GPU core on the M6 now has a built-in neural accelerator, and the chip carries two neural engines instead of one, which is the main reason local AI tasks got so much faster.
- In real testing, prompt processing on a 9-billion-parameter model jumped from around 210 tokens per second on the M4 to roughly 740 tokens per second on the M6 using llama.cpp.
- Image generation with Flux dropped from over a minute and a half on the M4 to about 35 seconds on the M6 for the same prompt and model.
- Xcode compile times for a 76-library Cocoa test project fell from 172 seconds on the M4 to 114 seconds on the M6, putting it close to an M1 Ultra Mac Studio.
- The M6 is Apple’s first 2nm chip (versus 3nm on the M4), and despite doing more work, it drew less peak power and ran a few degrees cooler under sustained GPU load.
- Thunderbolt 5 didn’t make it into the base M6 Mac mini; Apple reserved that for the Pro model, so heavy external storage or eGPU users won’t see a bandwidth jump there.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
What actually changed inside the M6 chip?
The M6 isn’t a simple clock-speed bump. The CPU now blends three kinds of cores instead of two: two “super cores” (Apple’s new name for its fastest cores), four performance cores, and six efficiency cores. That’s a more complex scheduling setup than the M4’s straightforward performance-plus-efficiency layout, and it shows up directly in multi-threaded work like compiling and heavy Python scripts.
The GPU jumped from 10 cores on the M4 to 12 cores on the M6, and every one of those cores now includes a neural accelerator. That feature debuted in the M5 generation but skipped the M4 entirely, so this is the single biggest reason AI workloads feel different. Apple also doubled the neural engine count, going from one to two.
Beyond compute, the M6 picks up Apple’s new N1 wireless chip (Wi-Fi 7, Bluetooth 6) and a 10-gigabit Ethernet jack, double the previous mini’s networking speed. The one holdout: Thunderbolt stays at version 4 on the base M6 mini. Apple kept Thunderbolt 5 exclusive to the Pro model, which matters if you move large files or LLM weights over an external drive regularly.
Manufacturing-wise, the M6 is built on a 2nm process, down from 3nm on the M4. Smaller transistors leak less power and run cooler, which in practice means similar idle power draw between the two machines but a bigger performance ceiling once you push the chip.
How much faster is the M6 in everyday and dev benchmarks?
Geekbench numbers alone are messy right now because Geekbench 7 replaced version 6 with lower baseline scores across the board, making cross-generation comparisons confusing unless you rerun old chips on the new version. That’s why benchmarks that reflect real workloads matter more here.
In Speedometer, which measures how fast a browser feels across React, Angular, and vanilla JavaScript workloads, the M6 hit a notably high score, described as the highest result seen in this kind of test to date.
For Python, a Mandelbrot-set benchmark that uses every available CPU core finished in about 21.5 to 32.8 seconds range differences between the two machines, with the M6 finishing meaningfully faster thanks to having 12 active cores versus the M4’s 10. Power draw during that fully loaded CPU test sat around 40 watts on the M4 and about 48 watts on the M6, showing the extra performance does come with extra draw under max CPU load (though not under GPU-bound AI tasks, as covered below).
Xcode compilation, a realistic proxy for daily developer work, took 172 seconds on the M4 and 114 seconds on the M6 for a project bundling 76 Cocoa libraries and their dependencies. For context, an M5 Max finished the same project in 84 seconds and an M1 Ultra Mac Studio landed at 112 seconds, putting the base M6 mini in the same ballpark as a previous-generation Ultra chip.
SSD performance also improved, especially in random 4K read/write at high queue depth, which is the pattern that matters most for compiling code or shuffling large model files. Sequential speeds were fast on both machines, but the M6 pulled ahead clearly on random I/O.
Does the M6 make a real difference for local AI?
This is where the generational gap is largest. Local AI performance depends on two things: compute (how fast the GPU and neural engine process a prompt) and memory bandwidth (how fast the system streams data while generating tokens).
Apple’s stated memory bandwidth figures are 120 GB/s for the M4 versus 153 GB/s for the M6 with 16GB of RAM, or 170 GB/s if you configure the M6 mini with 32GB. Independent testing using the STREAM benchmark measured closer to 112 GB/s on the M4 and around 143 to 145 GB/s on the M6, both somewhat below Apple’s marketing numbers but still showing a clear real-world gap.
Running a 9-billion-parameter GGUF model (Q4_K quantization) through llama.cpp, prompt processing went from about 210 tokens per second on the M4 to roughly 740 tokens per second on the M6. Token generation rose from about 18 tokens per second to nearly 27. Re-testing through llama-bench with a server-based setup (closer to how these machines get used in practice) showed time-to-first-token dropping from 2.5 seconds on the M4 to about 721 milliseconds on the M6, a difference you can see just by watching two chat windows generate text side by side.
Image generation showed a similar pattern. Running the same prompt through the Flux.1 Schnell model, the M6 produced its image in about 35 seconds while the M4 took over a minute and a half, roughly a 2.7x difference, on a model that fits comfortably within 16GB of memory on both machines.
Is the M6 more power-hungry or more efficient?
Despite the higher performance ceiling, the M6 didn’t need more power to get there in sustained AI workloads. Running a looped local-model generation test at 100% GPU utilization, the M6 drew a maximum of 38 watts compared to 44 watts on the M4, while finishing the work faster. Thermal imaging during the same test showed the M6 running a couple of degrees cooler (around 33 to 36 degrees versus 37 to 38 on the M4). Idle power draw was essentially identical between the two machines, so the 2nm process advantage shows up specifically when the chip is under load, not at rest.
Is upgrading from the M4 to the M6 Mac mini worth it?
For anyone running local language models, generating images locally, or compiling large codebases regularly, the jump is substantial rather than incremental. Prompt processing more than tripled in testing, image generation time dropped to roughly a third, and compile times fell by about a third as well, all while using less peak power under AI workloads. For general desktop use, web browsing, office work, and media consumption, the M4 remains capable, and the practical difference will be far less noticeable outside of AI and development-heavy tasks. The clearest upgrade case is anyone whose work is bottlenecked by GPU compute or memory bandwidth today.
Frequently Asked Questions
What’s the biggest hardware difference between the M6 and M4 chips?
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
The M6 adds neural accelerators to every GPU core, doubles the neural engine count to two, moves to a 12-core GPU from the M4’s 10-core GPU, and introduces a three-tier CPU core design (super, performance, efficiency cores) built on a 2nm process versus the M4’s 3nm process.
Does the M6 Mac mini support Thunderbolt 5?
No. Apple kept Thunderbolt 4 on the base M6 Mac mini and reserved Thunderbolt 5 for the Pro model, so users moving very large files over external Thunderbolt storage won’t see a speed increase there.
How much faster is the M6 for running local AI models?
In testing with a 9-billion-parameter model, prompt processing rose from about 210 tokens per second on the M4 to around 740 tokens per second on the M6, and image generation with Flux dropped from about 94 seconds to roughly 35 seconds for the same prompt.
Does the M6 use more power than the M4?
Not under sustained GPU-heavy AI workloads. In a 100% GPU utilization test, the M6 peaked at 38 watts versus 44 watts on the M4, while also running a few degrees cooler and finishing the work faster.
Is memory bandwidth different depending on the M6 configuration?
Yes. Apple lists 153 GB/s for M6 Mac minis with 16GB of memory and 170 GB/s if configured with 32GB, compared to about 120 GB/s on the M4. Independent STREAM benchmark testing measured somewhat lower real-world figures on both chips.