← Back to community benchmarks
Qwen3.5-122B-MXFP4-FLOAT16
User commentused mlx_lm.convert -q --q-mode mxfp4 --dtype float16
Performance
64k
tokens
221.5
PP tok/s
13.2
TG tok/s
295853
TTFT (ms)
74.4
Peak mem (GB)
Hardware
Chip
M1 Ultra (64c)
Memory
128 GB
GPU Cores
64
Software
oMLX
v0.2.10
macOS
macOS 26.3.1
Context
65,536
Performance by Context Length
| Context | PP tok/s | TG tok/s | Peak Mem | |
|---|---|---|---|---|
| 1k | 443.9 | 40.5 | 63.5 GB | view |
| 4k | 502.0 | 38.5 | 64.1 GB | view |
| 8k | 500.4 | 35.1 | 64.8 GB | view |
| 16k | 473.4 | 28.1 | 66.1 GB | view |
| 32k | 358.0 | 20.5 | 69.0 GB | view |
| 64k | 221.5 | 13.2 | 74.4 GB | current |
Batching Results
| Batch Size | TG tok/s | Speedup |
|---|---|---|
| 1× | 40.5 | 1.00× |
| 2× | 57.0 | 1.41× |
| 4× | 67.2 | 1.66× |