Research

Apple Silicon LLM Benchmarks: M1 to M5

Generation-by-generation local LLM throughput across Apple Silicon from M1 to M5. Memory bandwidth, decode speed by model size, the M5 Neural Accelerator step change, and which generation is worth upgrading from.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: July 2026

Five generations of Apple Silicon are still in active use for local inference. This page tracks what each generation delivers, where the real discontinuities are, and whether upgrading is worth it from where you are.

Memory Bandwidth by Generation

Bandwidth is the strongest single predictor of decode speed on Apple Silicon, so it is the right place to start.

ChipApproximate memory bandwidthMaximum unified memory
M1 / M1 Pro68-200 GB/s16-32 GB
M1 Max / Ultra400-800 GB/s64-128 GB
M2 / M2 Pro100-200 GB/s24-32 GB
M2 Max / Ultra400-800 GB/s96-192 GB
M3 / M3 Pro100-150 GB/s24-36 GB
M3 Max300-400 GB/s128 GB
M4 / M4 Pro120-273 GB/s32-64 GB
M4 Max~410-546 GB/s128 GB
M5 Max~546 GB/s128 GB

Note the M3 Pro regression against M2 Pro, which is real and caught a lot of buyers out. Note also that a base M3 has less bandwidth than an M1 Max from two generations earlier. Tier matters more than generation on Apple Silicon, and it always has.

Decode Throughput, 7B Class at Q4

ChipApproximate tokens per second
M1~20-30
M1 Max~45-60
M2 Max~55-70
M3 Max~60-75
M4 Max~75-90
M5 Max~95-110

Roughly a 4x improvement from M1 to M5 Max on the same workload, with the largest single-generation jump at M4 Max to M5 Max.

The M5 Discontinuity

M5 is the first genuine architectural change for AI work rather than an incremental bandwidth increase. Neural Accelerators added to the GPU cores give MLX a reported 4.06x time-to-first-token advantage over llama.cpp's Metal backend, which cannot currently reach them. On an M5 Max running a 35B-class sparse model in NVFP4, the accelerated path took prefill from 1,154 to 1,810 tokens per second and decode from 58 to 112.

The practical consequence is that M5's advantage over M4 is much larger on prompt-heavy agentic work than the raw bandwidth difference suggests, and it only materialises if you use MLX. See MLX versus llama.cpp.

Upgrade Guidance

Coming fromWorth upgrading to M5 Max?Reason
M1 / M2 baseYes, clearly3-4x throughput and far more memory
M1 Max / M2 MaxYes if you need capacityRoughly 1.6-2x throughput plus the accelerator path
M3 MaxMarginalRoughly 1.5x, hard to justify alone
M4 MaxOnly for prompt-heavy workModest decode gain, large first-token gain via MLX

Memory capacity, not chip generation, should drive the decision for most people. Moving from 32GB to 128GB changes which models you can run at all, which matters more than any throughput increase. That upgrade also got more expensive in 2026, see the memory shortage report.

Brand Visibility Implications

Apple Silicon put credible local inference on ordinary desks, and the installed base spans five generations. Even an M1 Max runs a 7B model fast enough for interactive use, which means unobservable brand answers have been generated on this hardware for years and the population capable of it keeps growing. See the local-LLM visibility blind spot.

Methodology

Vendor specifications come from NVIDIA, Apple, and model-card publications. Throughput figures aggregate community benchmark reporting from the llama.cpp discussions, the MLX repository, and published independent test suites. Single-stream decode unless stated otherwise. Ranges rather than point values are used wherever independent runs disagree, which is most of the time: quantisation format, prompt length, thermal state, and runtime version each move these numbers by more than the differences being measured. Treat every figure as an order-of-magnitude guide, not a specification. Updated quarterly.

How Presenc AI Helps

Presenc AI measures brand representation in the open-weight models that run on Apple Silicon, independent of chip generation or runtime.

Frequently Asked Questions

Coming from M1 or M2 base chips, clearly yes, with 3-4x the throughput and far more memory. From an M4 Max the decode gain is modest, but the M5 Neural Accelerators give MLX roughly a 4.06x time-to-first-token advantage, so it is worth it specifically for prompt-heavy agentic work. From an M3 Max the case is marginal.
Roughly 20-30 on M1, 45-60 on M1 Max, 55-70 on M2 Max, 60-75 on M3 Max, 75-90 on M4 Max, and 95-110 on M5 Max at Q4 quantisation. That is about a 4x improvement across the generations, with the largest single jump between M4 Max and M5 Max.
Apple reduced memory bandwidth on the M3 Pro relative to the M2 Pro. Because bandwidth is the strongest predictor of decode speed, that shows up directly as lower tokens per second. It is a real regression and a reminder that tier matters more than generation number on Apple Silicon.
Memory, for most people. Going from 32GB to 128GB changes which models you can run at all, while a chip generation typically changes how fast you run the models you could already run. Capacity is a hard constraint and speed is a comfort one.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.