Leaderboard
On-device LLM performance rankings powered by Glicko-2
iPad Pro 12.9 inch 7th Gen
iOSRank
#2
Rating
2,013
±16 RD
Win Rate
99.5%
Conservative Rating
1,980
TG Rating
2,011
PP Rating
2,013
Matches
1,005
Record
1000W – 5L
Models Tested
| Model | TG Median (tok/s) | PP Median (tok/s) | TG Best | PP Best | Runs |
|---|---|---|---|---|---|
| gemma-3-1b-it.Q4_K_S | 85.61 | 181.24 | 85.61 | 181.24 | 1 |
| llama-3.2-1b-instruct-q8_0 | 58.90 | 1350.52 | 58.90 | 1350.52 | 1 |
| gemma-2-2b-it-Q6_K | 40.96 | 589.76 | 42.40 | 663.20 | 3 |
| qwen2.5-3b-instruct-q5_k_m | 36.93 | 424.87 | 36.93 | 424.87 | 1 |
| Qwen3-4B-UD-Q4_K_XL | 34.96 | 346.05 | 34.96 | 346.05 | 1 |
| gemma-2-2b-it.Q5_K_M | 30.49 | 551.99 | 31.86 | 556.03 | 2 |
| Yi-Coder-1.5B-Chat.fp16 | 30.48 | 64.21 | 30.48 | 64.21 | 1 |
| Phi-3.5-mini-instruct.Q4_K_M | 30.05 | 196.02 | 35.98 | 358.14 | 2 |
| qwen2.5-3b-instruct-q8_0 | 28.99 | 62.93 | 28.99 | 62.93 | 1 |
| Llama-3.2-3B-Instruct-Q6_K | 28.38 | 376.20 | 34.43 | 479.35 | 6 |
| Qwen3-8B.Q3_K_M | 18.97 | 166.12 | 18.97 | 166.12 | 1 |
| ai21labs_AI21-Jamba-Reasoning-3B-Q8_0 | 17.78 | 43.16 | 17.78 | 43.16 | 1 |
| DeepSeek-R1-0528-Qwen3-8B-IQ4_NL | 11.22 | 15.65 | 11.22 | 15.65 | 1 |
| DeepSeek-R1-Distill-Qwen-7B-Q4_K_L | 2.41 | 10.63 | 2.41 | 10.63 | 1 |
| Llama-3.2-9B-Uncensored-Brainstorm-Alpha-D_AU-IQ4_XS | 0.23 | 13.14 | 0.23 | 13.14 | 1 |
Head-to-Head Record
1–50 of 268 rows
1 / 6
Performance by App Version
ImprovedRegressed