M3 LLM benchmark
No signed M3 submissions yet. Check the configuration guidance below, or contribute a reproducible measurement from your own machine.
Specs vendor specs, not measured
- Unified memory (max configuration)
- 24 GB
Plan memory before choosing a model
The catalog lists a family maximum of 24 GB of unified memory. Enter your installed amount; smaller configurations have less space. Model weights, context cache and runtime overhead must share the available memory. Capacity alone does not establish which model will fit or how fast it will run. Use the memory-fit calculator for your model and context as an estimate, then verify with the intended runtime. No signed M3 measurement is available yet.
Contribute a signed M3 benchmark
Run on YOUR hardware to populate this page: pipx install llm-speed && llm-speed bench
$ pipx install llm-speed && llm-speed bench
Community folklore on M3
7 unverified claims extracted from Reddit/HN comments. Lower trust than signed runs above; every row links to the source.
- communityconfidence 70%
40.00tok/s — Phi-4 Mini on M3 via llama.cpp
our signed data: M3 · Phi-4 Mini
“t/s). Every rival model loses 70% speed beyond 32k **High intelligence per token ratio:** * Scored 0.31 combined intelligence index at \~40 t/s, above Gemma 3 4B (0.20) and Phi-4 Mini (0.22) * Qwen 3 4B ranks slightly higher in raw score (0.35) but runs 3x slower **Outpaces IB…”
- communityconfidence 60%
86.00tok/s — Qwen3-30B-A3B on M3 via mlx
our signed data: M3 · Qwen3-30B-A3B
“w/ MLX: 2,320 t/s 🥈 3090: 2,157 t/s 🥉 M3 w/ Metal: 1,614 t/s tg128: 🥇 3090: 136 t/s 🥈 M3 w/ MLX: 97 t/s 🥉 M3 w/ Metal: 86 t/s https://preview.redd.it/7f1bj2ag4rkf1.png?width=2522&format=png&auto=webp&s=44184256681e46b0bb2d8324c4d6abda8b7f4266”
- communityconfidence 60%
97.00tok/s — Qwen3-30B-A3B on M3 via mlx
our signed data: M3 · Qwen3-30B-A3B
“results! pp512: 🥇M3 w/ MLX: 2,320 t/s 🥈 3090: 2,157 t/s 🥉 M3 w/ Metal: 1,614 t/s tg128: 🥇 3090: 136 t/s 🥈 M3 w/ MLX: 97 t/s 🥉 M3 w/ Metal: 86 t/s https://preview.redd.it/7f1bj2ag4rkf1.png?width=2522&format=png&auto=webp&s=44184256681e46b0bb2d8324c4d6abda…”
- communityconfidence 55%
15.00tok/s — on M3 INT4
our signed data: M3
“nt," it specializes in deep reasoning, autonomous tool orchestration, and coding. After running it on my setup with two M3 Ultras at around 15 tokens per second, I can vouch for its efficiency and capabilities. The 256K context window handled large projects without hiccups, and i…”
- communityconfidence 50%
1.00tok/s — Llama 3.2 3B on M3
our signed data: M3 · Llama 3.2 3B
“le:** * Keeps producing \~40 tokens per second on Mac even past 32k context * Still cranks out \~33 t/s at 128k while Qwen 3 4B drops to <1 t/s and Llama 3.2 3B goes down to \~5 t/s **Best long context efficiency:** * From 1k to 128k context, latency barely moves (43 to 33 t/…”
- communityconfidence 50%
33.00tok/s — Llama 3.2 3B on M3
our signed data: M3 · Llama 3.2 3B
“y findings: **Faster and steadier at scale:** * Keeps producing \~40 tokens per second on Mac even past 32k context * Still cranks out \~33 t/s at 128k while Qwen 3 4B drops to <1 t/s and Llama 3.2 3B goes down to \~5 t/s **Best long context efficiency:** * From 1k to 128k c…”
- communityconfidence 50%
40.00tok/s — Llama 3.2 3B on M3
our signed data: M3 · Llama 3.2 3B
“following: * Mac M3 36GB * iPhone 16 Pro * Galaxy S25 Here are our key findings: **Faster and steadier at scale:** * Keeps producing \~40 tokens per second on Mac even past 32k context * Still cranks out \~33 t/s at 128k while Qwen 3 4B drops to <1 t/s and Llama 3.2 3B goes …”
Request a run on M3
No measured run yet for these pairs. Vote for the one you want; the most-requested runs go first. See all requests.
We never sell or share your email. Privacy