Is the M4 faster than the M3 for LLMs?

The M4 reads memory about 20% faster than the M3 (120 vs 100 GB/s), so it writes about 20% faster with any model that fits both. Reading a long prompt depends on GPU compute instead, which this comparison doesn’t cover.

Which can run bigger models, the M3 or the M4?

The M4: The M4 goes up to 32 GB of memory, the M3 to 24 GB. At its largest it runs 27 of 73 open models well, against 20.

M3 vs M4 for local LLMs: memory, speed, which to buy

Side by side

	M3	M4
Memory sizes	8, 16, 24 GB	16, 24, 32 GB
Memory bandwidth	100 GB/s	120 GB/s
For models, at most	about 18 GB	about 24 GB
Models that run well	20 of 73	27 of 73
MLX seat class	Core	Core
Found in	MacBook Pro (14-inch, M3, Nov 2023), iMac (24-inch, M3, 2023, Two ports), iMac (24-inch, M3, 2023, Four ports), MacBook Air (13-inch, M3, 2024), MacBook Air (15-inch, M3, 2024)	MacBook Pro (14-inch, M4, 2024), iMac (24-inch, 2024, Two ports), iMac (24-inch, 2024, Four ports), Mac mini (2024), MacBook Air (13-inch, M4, 2025), MacBook Air (15-inch, M4, 2025)

Speed with popular models

Tokens a second for one conversation, each chip at its largest memory size (24 GB and 32 GB), with the best version of the model that fits.

Model	M3	M4
Llama 3.1 8B Instruct8B	10–18 4-bit	12–21 4-bit
gpt-oss-20b21B · 3.6B active	24–42 native 4.25-bit	28–51 native 4.25-bit
Qwen3.6 35B A3B35B · 3B active	35–62 3-bit lower quality	32–56 4-bit
Qwen3.8 27B27B	3–5 4-bit	2–4 6-bit
Llama 3.3 70B Instruct70B	too big	2–4 2-bit lower quality
gpt-oss-120b117B · 5.1B active	too big	too big
Qwen3.5 122B A10B122B · 10B active	too big	too big
DeepSeek V4 Flash 0731284B · 13B active	too big	too big

Questions

Is the M4 faster than the M3 for LLMs?: The M4 reads memory about 20% faster than the M3 (120 vs 100 GB/s), so it writes about 20% faster with any model that fits both. Reading a long prompt depends on GPU compute instead, which this comparison doesn’t cover.
Which can run bigger models, the M3 or the M4?: The M4: The M4 goes up to 32 GB of memory, the M3 to 24 GB. At its largest it runs 27 of 73 open models well, against 20.

More comparisons

The M3 for local AI · The M4 for local AI · Compare every model · The best Mac for local LLMs

How it’s worked out

Memory: the model’s parameters at 8, 6, 4, 3 or 2 bits per weight plus MLX’s group scales (half a bit per weight), or its native format, with room for the context on top. By default macOS lets the GPU use about 75% of unified memory. Speed: memory bandwidth divided by the bytes read per token is the ceiling for one conversation; in llama.cpp’s public Apple Silicon benchmarks real runs reach about 45% to 80% of it, which is the range shown. These are estimates, not measurements: measure your own Mac on MLX Bench. Model figures from each official model card, chip figures from Apple’s Tech Specs. Checked 2026-10-19. Same data as mlx.app/compare.