Performance Benchmarks
Explore momo-kibidango's performance across different Apple Silicon chips and configurations.
Performance Comparison
Baseline (Sonnet 3.5 only)12.5 tok/s
2-Model Speculative18.7 tok/s
Pyramid (momo-kibidango)24.6 tok/s
vs Baseline
+97x
vs 2-Model
+32x
Speedup
1.97x
Acceptance Rates
89%
Tier 1 → Tier 2
Haiku 2 → Haiku 3
76%
Tier 2 → Tier 3
Haiku 3 → Sonnet 3.5
68%
Overall
End-to-end acceptance
Higher acceptance rates mean fewer rejections and faster overall inference.
Memory Usage
Model Sizes
Haiku 2 (Tier 1)2.3 GB
Haiku 3 (Tier 2)3.8 GB
Sonnet 3.5 (Tier 3)4.9 GB
Total11.6 GB
Runtime Requirements
Minimum RAM16 GB
Recommended RAM32 GB
Peak Usage13.2 GB
Disk Space50 GB
Export Benchmark Data
Download the complete benchmark dataset for your own analysis.