M4 LLM benchmark
No signed M4 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)
- 32 GB
Plan memory before choosing a model
The catalog lists a family maximum of 32 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 M4 measurement is available yet.
Contribute a signed M4 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 M4
12 unverified claims extracted from Reddit/HN comments. Lower trust than signed runs above; every row links to the source.
- communityconfidence 60%
180.0tok/s — Qwen3-32B on M4 via lm-studio
our signed data: M4 · Qwen3-32B
“its 79.51% while sustaining ~64 tok/s - arguably today's best speed/quality trade-off for Mac setups. 5️⃣ The 0.6B micro-model races above 180 tok/s but tops out at 37.56% - that's why it's not even on the graph (50 % performance cut-off). All local runs were done with @lmstudi…”
- communityconfidence 60%
180.0tok/s — Qwen3-32B on M4 via lm-studio
our signed data: M4 · Qwen3-32B
“9.51%** while sustaining \~64 tok/s - arguably today's best speed/quality trade-off for Mac setups. 5. The **0.6B** micro-model races above 180 tok/s but tops out at **37.56%** \- that's why it's not even on the graph (50 % performance cut-off). All local runs were done with LM …”
- communityconfidence 60%
35.64tok/s — qwen2.5-coder on M4 via ollama
our signed data: M4 · qwen2.5-coder
“ATE: The issue was resolved by killing the processes that kept VRAM occupied, and then Ollama with the qwen2.5-coder:32b model performed at 35.64 tokens/s. I guess I won't need to spend money on M4 for now :)”
- communityconfidence 55%
39.00tok/s — on M4 via mlx fp16
our signed data: M4
“here are some benchmarks for gemma-3-text-1B-it-mlx on a base m4 mbp: q3 - 125 tps q4 - 110 tps q6 - 86 tps q8 - 66 tps fp16 I think - 39 tps Edit: to be clear the models that now are working are called **alexgusevski/gemma-3-text-...** or **mlx-community/gemma-3-text-...**…”
- communityconfidence 55%
66.00tok/s — on M4 via mlx fp16
our signed data: M4
“e -text. Just for fun here are some benchmarks for gemma-3-text-1B-it-mlx on a base m4 mbp: q3 - 125 tps q4 - 110 tps q6 - 86 tps q8 - 66 tps fp16 I think - 39 tps Edit: to be clear the models that now are working are called **alexgusevski/gemma-3-text-...** or **mlx-commu…”
- communityconfidence 55%
86.00tok/s — on M4 via mlx fp16
our signed data: M4
“ar, notice the -text. Just for fun here are some benchmarks for gemma-3-text-1B-it-mlx on a base m4 mbp: q3 - 125 tps q4 - 110 tps q6 - 86 tps q8 - 66 tps fp16 I think - 39 tps Edit: to be clear the models that now are working are called **alexgusevski/gemma-3-text-...** o…”
- communityconfidence 55%
110.0tok/s — on M4 via mlx fp16
our signed data: M4
“t-mlx or similar, notice the -text. Just for fun here are some benchmarks for gemma-3-text-1B-it-mlx on a base m4 mbp: q3 - 125 tps q4 - 110 tps q6 - 86 tps q8 - 66 tps fp16 I think - 39 tps Edit: to be clear the models that now are working are called **alexgusevski/gemma-…”
- communityconfidence 55%
125.0tok/s — on M4 via mlx fp16
our signed data: M4
“-3-text-1B-8bit-mlx or similar, notice the -text. Just for fun here are some benchmarks for gemma-3-text-1B-it-mlx on a base m4 mbp: q3 - 125 tps q4 - 110 tps q6 - 86 tps q8 - 66 tps fp16 I think - 39 tps Edit: to be clear the models that now are working are called **alexg…”
- communityconfidence 55%
85.00tok/s — on M4 via mlx BF16
our signed data: M4
“VL 4B and 30B on LMStudio yesterday, so there's other frameworks that support it too. As a single point of comparison, my M4 macbook gets 85 tokens/second interpreting an image with Qwen3 VL 4B MLX 4bit, so about 1/3 as fast - but also probably a significantly smaller quant than…”
- communityconfidence 55%
42.00tok/s — on M4 FP4
our signed data: M4
“Great setup! Funny that I get 42t/s on that 35B FP4 with 100k context on a MacBook M4 with 32GB. I thought you could get much more with these 256GB of yours.”
Request a run on M4
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