Benchmarks

Every number below is a modelled estimate from memory bandwidth and quantized weight size — we will not print a measured figure we did not receive. Real runs from real Macs replace estimates as they arrive.

Ran a model on your Mac and timed it? Submit the run — chip, memory, model, quantization, runtime version and tokens per second. Submissions are published as unverified until several independent runs agree.

Estimated generation speed, tokens per second, best quantization that fits at your 48 GB and current context
Chip	GB/s	Llama 3.2 3B Instruct	Qwen3 4B	Mistral 7B Instruct v0.3	Qwen2.5 7B Instruct	InternVL3 8B
M1	68	7.6	11.5	6.3	3.2	5.8
M1 Pro	200	23.7	35.8	19.6	10.0	17.9
M1 Max	400	49.2	74.4	40.7	20.8	37.2
M1 Ultra	800	101	152	83.5	42.6	76.2
M2	100	11.4	17.2	9.4	4.8	8.6
M2 Pro	200	24.0	36.2	19.9	10.1	18.1
M2 Max	400	49.8	75.3	41.3	21.1	37.6
M2 Ultra	800	102	154	84.6	43.2	77.2
M3	100	11.4	17.2	9.4	4.8	8.6
M3 Pro	150	17.5	26.5	14.5	7.4	13.2
M3 Max (14-core CPU)	300	36.4	55.1	30.2	15.4	27.5
M3 Max (16-core CPU)	400	49.8	75.3	41.3	21.1	37.6
M3 Ultra	819	106	160	87.6	44.7	80.0
M4	120	13.8	20.9	11.4	5.8	10.4
M4 Pro	273	33.2	50.1	27.5	14.0	25.1
M4 Max (14-core CPU)	410	51.7	78.1	42.8	21.8	39.1
M4 Max (16-core CPU)	546	69.7	105	57.7	29.5	52.7
M5	153	17.9	27.0	14.8	7.5	13.5
M5 Prounverified	307	37.8	57.1	31.3	16.0	28.5
M5 Maxunverified	614	79.4	120	65.7	33.5	60.0

All figures marked est. — estimates derived from published bandwidth, assuming a warm model, short prompt and no thermal throttling. Expect real Macs to land within roughly ±25%.

Measured runs from real Macs

Loading…

How we would verify a submission

Same model repo, same quantization, same runtime version.
Warm start: the model already resident, so the first load is not counted.
At least 256 generated tokens, reported as the median of three runs.
No other heavy applications open, and the machine on mains power.
Three independent submissions within 15% of each other before a number stops being labelled unverified.