M5 MacBook Air: How Much Faster Is Local AI, Really?
The M5 MacBook Air brings a real memory bandwidth jump for local LLMs. Here's what that means for prompt processing and token generation.
Drafted with Claude from source material, checked by automated verification, and reviewed before release. How we make these.
What does the M5 change for local AI on a MacBook Air?
The M5 MacBook Air raises memory bandwidth to 142 GB per second, up 26 percent from the M4’s 113 GB per second, based on the STREAM benchmark. That number matters more than raw CPU speed for running local large language models, because once a model is loaded, token generation is bottlenecked by how fast the chip can move data between memory and the GPU, not by how fast the cores crunch instructions. Apple’s own claim of a 28 percent bandwidth increase lines up almost exactly with the measured result, which is a rare case of marketing numbers holding up under an actual benchmark run.
TL;DR
- The M5’s memory bandwidth hit 142 GB/s in STREAM testing, a 26 percent jump over the M4’s 113 GB/s, closely matching Apple’s own 28 percent claim.
- Local LLM performance splits into two phases: prompt processing, which depends on raw compute, and token generation, which depends almost entirely on memory bandwidth.
- Across five MacBook Air generations (M1 through M5), general CPU gains have slowed year over year, with the M5’s Speedometer score up only about 5 percent over the M4, the smallest jump since the M1.
- Multi-core and sustained workloads tell a different story: a Llama.cpp build was about 14 percent faster on the M5 than the M4, and nearly twice as fast as the original M1.
- The M5 keeps the same 10-core CPU layout as the M4, but Apple now calls four of those cores “super cores,” and the architecture change adds real gains even without more cores.
- Storage on the M5 got a dramatic upgrade, with sequential read and write speeds more than doubling prior generations, which matters when downloading and swapping large model weight files.
- Sustained, all-core workloads still cause thermal throttling on the M5’s fanless chassis, with the biggest performance drop of any generation tested (about 6 percent from first to last iteration in a five-round stress test).
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Why does memory bandwidth matter more than CPU speed for LLMs?
Running a large language model locally involves two distinct stages, and they stress different parts of the chip. The first stage, prompt processing, reads and encodes the input you give the model. This is compute-heavy and benefits from more cores, wider execution units, and dedicated matrix acceleration hardware. The second stage, token generation, produces the model’s output one token at a time. Each step requires pulling the model’s weights back through memory, and for a model with billions of parameters, that means moving gigabytes of data per second. If memory bandwidth is the bottleneck, adding CPU cores does very little to speed up generation. That is why Apple Silicon chips built for AI work are judged as much on their bandwidth numbers as their core counts.
This is also why unified memory architecture matters so much for on-device AI. Because the CPU, GPU, and neural accelerators all share the same pool of RAM, a bandwidth increase benefits every part of the local AI pipeline at once, not just one component.
How does the M5 compare across five generations of MacBook Air?
Looking at the last five years of Apple Silicon Airs side by side shows a pattern: generational CPU gains have been shrinking, while multi-core and system-level throughput keeps climbing.
On Speedometer, a browser benchmark that reflects real-world responsiveness for web apps and Electron-based editors like VS Code, the jumps were 22 percent (M1 to M2), 14 percent (M2 to M3), 11 percent (M3 to M4), and just 5 percent (M4 to M5). That’s the smallest year-over-year gain in the lineup, though the cumulative M1-to-M5 improvement is around 60 percent.
Single-core compiled workloads tell a similar story. A single-threaded C++ sort test dropped from 2 minutes 38 seconds on the M1 to 2 minutes 19 seconds on the M5, an 8 percent gain over the M4 and a 36 percent cut from the M1, even though core count didn’t change between the two latest chips.
Multi-core and real build workloads show the bigger wins. Compiling Llama.cpp from source took 1 minute 40 seconds on the M1 and just 54 seconds on the M5, a 14 percent improvement over the M4 and nearly half the original time. The jump from 8 cores (M1 through M3) to 10 cores (M4 and M5) explains part of this, along with Apple’s new “super core” design on four of the M5’s ten cores.
Does the M5 throttle under sustained AI-scale workloads?
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Yes, and more than any previous MacBook Air tested. In a sustained stress test using a merge sort across a billion integers on all cores, repeated over five iterations, the M5 finished fastest overall at 129 seconds per iteration by the end, but showed the largest performance drop from first to last iteration: about 6 percent. The M4 and M3 each dropped about 5 percent. The M2 was the most stable, losing only 3 percent, though it was also the slowest machine overall. The M1, run passively with no fan, stayed noticeably cooler to the touch throughout.
This matters for anyone planning to run local LLMs for extended sessions, batch inference jobs, or long fine-tuning-adjacent tasks on a MacBook Air. The fanless chassis means sustained, all-core AI workloads will eventually hit a thermal ceiling and slow down, even if the initial burst speed is the best Apple has shipped in this form factor. A MacBook Pro with active cooling would not show the same drop under identical load.
Is the M5 MacBook Air worth it for running local LLMs?
For anyone using local models like Qwen through frameworks such as MLX, the M5’s bandwidth increase is the headline change. Faster memory bandwidth means noticeably quicker token generation on the same model size compared to the M4, and a large jump compared to the M1 through M3. Prompt processing also benefits from the M5’s neural accelerators, which are built into the GPU cores and designed to speed up the matrix-heavy math involved in reading and encoding long prompts.
Storage is the other quiet upgrade that matters for local AI work. The M5 more than doubles sequential read and write speeds compared to earlier models, and Apple dropped the 256 GB base configuration in favor of starting higher and offering options up to 4 TB. Anyone who has tried to keep multiple quantized model files, datasets, and development tools on a MacBook Air will recognize why that matters. Storage speed also affects how quickly large model weights load from disk into memory in the first place.
The catch is memory capacity, not bandwidth. The M1 through M3 shipped with 8 GB as a base configuration, while the M4 and M5 start at 16 GB. Larger local models need more memory just to fit, regardless of how fast that memory moves data, so anyone planning to run bigger quantized models locally should treat memory size as a separate, and arguably more important, decision than which chip generation to buy.
Frequently Asked Questions
What is the actual memory bandwidth of the M5 chip?
The M5 measured 142 GB per second in STREAM benchmark testing, compared to 113 GB per second on the M4, a 26 percent increase that closely matches Apple’s stated 28 percent improvement.
Does more CPU cores help with local LLM performance?
Core count mainly helps with prompt processing and general compute tasks. Token generation, the slower and more visible part of running a local LLM, depends primarily on memory bandwidth, not core count.
How much faster is the M5 than the M4 for AI workloads?
Based on memory bandwidth alone, the M5 is about 26 percent faster than the M4. Real-world token generation gains will track close to that figure since generation speed is bandwidth-bound.
Does the MacBook Air throttle when running heavy local AI workloads?
Yes. Sustained, all-core workloads cause measurable slowdowns on every Air generation tested, and the M5 showed the largest drop, around 6 percent, from its first workload iteration to its last.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
Should I buy 16 GB or more memory for local AI on the M5?
Since the M4 and M5 both start at 16 GB minimum, anyone planning to run larger local models should consider upgrading further, since memory capacity determines what size model can be loaded at all, separate from how fast that memory moves data.