← Back to community benchmarks
Gemma-4-21B-REAP-JANG3M
Performance
16k
tokens
1,710
PP tok/s
77.7
TG tok/s
9580
TTFT (ms)
11.5
Peak mem (GB)
Hardware
Chip
M3 Ultra (60c)
Memory
96 GB
GPU Cores
60
Software
oMLX
v0.3.12
macOS
macOS 26.5
Context
16,384
Performance by Context Length
| Context | PP tok/s | TG tok/s | Peak Mem | |
|---|---|---|---|---|
| 1k | 1,339 | 89.5 | 10.2 GB | view |
| 4k | 1,735 | 88.0 | 10.9 GB | view |
| 8k | 1,793 | 82.6 | 11.0 GB | view |
| 16k | 1,710 | 77.7 | 11.5 GB | current |
| 32k | 1,518 | 66.7 | 12.4 GB | view |
| 64k | 1,220 | 47.0 | 14.5 GB | view |
| 128k | 852.4 | 29.7 | 18.6 GB | view |
| 195k | 626.2 | 22.2 | 23.6 GB | view |
Batching Results
| Batch Size | TG tok/s | Speedup |
|---|---|---|
| 1× | 89.5 | 1.00× |
| 2× | 133.4 | 1.49× |
| 4× | 199.8 | 2.23× |
| 8× | 249.3 | 2.79× |