Benchmark Dashboard

Live benchmark data from CI. Updated after each successful perf run.

Decode Speed After Compaction

Model	Context	Ratio	Cosine	Compacted (t/s)	Baseline (t/s)	Speedup
Qwen3-8B	8K	8x	0.997	9.3	5.7	+63%
Qwen3-30B-A3B	4K	8x	0.999	20.2	13.5	+50%
Qwen3-8B	4K	8x	0.997	13.8	10.7	+29%
Qwen3-8B	16K	8x	0.999	5.3	4.8	+10%

Compaction Quality Heatmap

Logit cosine similarity (select pipeline). Higher is better. Green >= 0.99, yellow >= 0.95.

Model	4K/2x	4K/4x	4K/8x	8K/2x	8K/4x	8K/8x
Qwen3-8B	0.999	0.999	0.997	0.999	0.998	0.997
Qwen3-30B-A3B	0.999	0.999	0.999	-	-	-
Qwen3-14B	0.995	0.996	0.992	0.999	0.997	0.973
DeepSeek-R1-14B	0.999	0.998	0.993	-	-	-

Fork vs Upstream vs Ollama (Baseline)

No compaction — pure decode speed comparison. Zero regression from compaction code.

256K Iterative Context Extension

64K physical KV cache, 49-57 compaction cycles, 256K effective tokens.

Platform Build Status

Platform	CI Workflow	Status
macOS (Apple Silicon)	modelai-ci	Active
macOS (Server smoke)	modelai-server-smoke	Active
Windows (MSVC x64)	modelai-ci-windows	Build-only
Perf regression	modelai-perf-smoke	Active
Upstream sync	modelai-upstream-sync	Weekly

Upstream Sync Health