MLXHub · Benchmark v1
Device ranking
Global score in Benchmark v1. 1000 = iPad Pro 12.9″ (M1) with 8 GB, so a score of 1570 means about 1.57 times the performance of that reference.
- Data updated October 11, 2026
- Devices ranked: 2
- Best result per device model and memory
- What does 1000 mean?
-
Profile compared with the reference
Each metric as a multiple of the iPad Pro 12.9″ (M1), 8 GB. Bars start at 0; the line marks 1×.
- Generation 1.39×
- Usable memory 0.75×
- Cold prompt processing 5.20×
- Endurance 1.06×
- Expert reading 2.61×
- Model loading 2.18×
Below the reference: usable memory, 4.39 GB compared with 5.84 GB on the M1.
- 4-bit models
- 1,352.91
- 8-bit models
- 1,654.93
Score by category
Generation speed
44.5 tok/s
Gemma 4 E2B 4-bit · M1: 36.6 tok/s
Usable memory
4.39 GB
M1: 5.84 GB
Measured with
iOS 27.0 · MLXHub 2.2.1
Measured on Sep 16, 2026
Under the hood
Tokens per second while each scenario generates, about one sample per second, with the device's thermal state behind. Breaks in the line are model loading and prompt processing.
Gemma 4 E2B4-bitNominal Average 43.5 tok/sBattery −5 %
Gemma 4 E4B4-bitFair Average 22.1 tok/sBattery −5 %
Gemma 4 E2B8-bitNominal Average 24.8 tok/sBattery −5 %
LFM2.5 8B A1B8-bitExperts from storageSerious Average 5.2 tok/sBattery −10 %83 s serious or worse
Every scenario, as measured Scenario Prompt tokens Generated Prefill tok/s Decode tok/s Load GB/s Endurance Storage MB/s Gemma 4 E2B 4-bit short 421 5,696 917.4 44.5 2.27 0.963 — Gemma 4 E2B 4-bit long 4,541 3,232 3,307.2 41.8 2.07 0.982 — Gemma 4 E4B 4-bit short 421 2,624 426.2 22.1 2.17 0.927 — Gemma 4 E4B 4-bit long 4,541 1,056 933.2 21.9 3.11 1.000 — Gemma 4 E2B 8-bit short 421 3,872 257.4 24.9 3.25 0.981 — Gemma 4 E2B 8-bit long 4,541 2,432 3,299.3 24.4 3.25 0.994 — LFM2.5 8B A1B 8-bit short 421 608 66.6 5.1 9.73 0.937 3,397 LFM2.5 8B A1B 8-bit long 4,541 96 83.4 5.1 9.62 0.776 3,016 - Usable memory
- 4.39 GB
- Peak footprint
- 5.17 GB
- Physical memory
- 12.3 GB
- Test duration
- 19 s
- Stopped at
- Critical memory warning
Memory test
- Total duration
- 38.8 min
- CPU cores
- 6
- OS build
- 24A437
- App build
- 84
- Battery
- 80 % → 55 %
- Low Power Mode
- No
- Charging
- No
- Peak thermal state
- Serious
Run
-
Profile compared with the reference
Each metric as a multiple of the iPad Pro 12.9″ (M1), 8 GB. Bars start at 0; the line marks 1×.
- Generation 1.00×
- Usable memory 1.00×
- Cold prompt processing 1.00×
- Endurance 1.00×
- Expert reading 1.00×
- Model loading 1.00×
This is the reference device, so every metric is exactly 1×.
- 4-bit models
- 1,000.00
- 8-bit models
- 1,000.00
Score by category
Generation speed
36.6 tok/s
Gemma 4 E2B 4-bit · M1: 36.6 tok/s
Usable memory
5.84 GB
M1: 5.84 GB
Measured with
iPadOS 27.0 · MLXHub 2.2.1
Measured on Sep 16, 2026
Under the hood
Tokens per second while each scenario generates, about one sample per second, with the device's thermal state behind. Breaks in the line are model loading and prompt processing.
Gemma 4 E2B4-bitNominal Average 35.5 tok/sBattery −5 %
Gemma 4 E4B4-bitNominal Average 18.7 tok/sBattery −5 %
Gemma 4 E2B8-bitSerious Average 15.4 tok/sBattery −10 %210 s serious or worse
LFM2.5 8B A1B8-bitExperts from storageSerious Average 3.7 tok/sBattery −10 %177 s serious or worse
Every scenario, as measured Scenario Prompt tokens Generated Prefill tok/s Decode tok/s Load GB/s Endurance Storage MB/s Gemma 4 E2B 4-bit short 421 2,304 207.9 36.6 1.21 0.965 — Gemma 4 E2B 4-bit long 4,541 928 671.3 32.4 1.18 1.000 — Gemma 4 E4B 4-bit short 421 928 73.8 18.8 1.01 0.961 — Gemma 4 E4B 4-bit long 4,541 224 162.7 18.2 1.09 1.000 — Gemma 4 E2B 8-bit short 421 928 15.4 13.5 0.99 0.786 — Gemma 4 E2B 8-bit long 4,541 544 243.6 13.0 1.11 0.676 — LFM2.5 8B A1B 8-bit short 421 288 33.1 3.7 5.65 1.000 1,106 LFM2.5 8B A1B 8-bit long 4,541 64 51.7 3.9 6.37 0.771 1,366 - Usable memory
- 5.84 GB
- Peak footprint
- 6.87 GB
- Physical memory
- 7.9 GB
- Test duration
- 92 s
- Stopped at
- Critical memory warning
Memory test
- Total duration
- 48.2 min
- CPU cores
- 8
- OS build
- 24A435
- App build
- 84
- Battery
- 65 % → 35 %
- Low Power Mode
- No
- Charging
- No
- Peak thermal state
- Serious
Run
No results have been published for this device type yet.
Only a few devices have been published so far. Run the benchmark in MLXHub and publish your result to add yours.
About the results
How to read the ranking
Result labels
Published and checked by the app, not yet reviewed. Included in the ranking.
Passed the automated consistency review.
Verified by the MLXHub team.
Weight of each metric in the global score
- 30%Generation
- 22%Usable memory
- 15%Cold prompt processing
- 15%Endurance
- 10%Expert reading
- 8%Model loading
Results are published by MLXHub users; the MLXHub team verifies them periodically, both manually and through automated reviews. Each row shows the best result published for one device model and memory size.
How the benchmark worksRanking data
The ranking is published as a snapshot of results that passed the app's checks and were not removed by review. Scores are only comparable within the same benchmark version.
Data updated October 11, 2026 · Benchmark v1