Inference throughput
Live performance
Testing
0.0
tok/s
Discovering task
0%
Time to first token
— ms
Memory peak
— GB
Efficiency
— tok/W
Auto-discovered task
MLX · Metal
Llama 3.2 · 3B Instruct
Optimized decode benchmark selected automatically for this device's unified-memory profile.
Precision
4-bit quantized
Context
4,096 tokens
Batch
1 × 256 output
Backend
MLX fused kernels
Rolling sample
Throughput stability
±0.0%
Start
Live samples
Now