context-compressor benchmark — 2026-08-24

Model: llama-3.2-8x3b-moe-dark-champion-instruct-uncensored-abliterated-18.4b

Variant: compact

Settings: {"limitTokens":500,"floorTokens":400,"keepRecentTokens":300,"chunkTokens":800,"summarizerMaxTokens":400,"consolidationBudgetTokens":600,"consolidationKeepNewest":2,"mergeMaxTokens":400,"maxExcerptChars":32000,"quizMaxTokens":120}

Fixture	Recall (compacted)	Recall (consolidated)	Reduction	Chunk latency mean/p50/max (ms)	Structure	Cache
coding-agent	5/5	—	56%	990.3/1113/1338	✅	✅
research	1/3	—	46%	1014/1014/1014	✅	✅
contradictory-updates	3/3	—	36%	469/469/469	✅	✅
tool-spam	3/3	—	70%	597.8/497/1063	✅	✅

Recall = seeded facts answered correctly from the compacted view alone. Results are comparable only within one model; see bench/README.md.

context-compressor benchmark — 2026-08-24

Model: llama-3.2-8x3b-moe-dark-champion-instruct-uncensored-abliterated-18.4b

Variant: compact

Fixture	Recall (compacted)	Recall (consolidated)	Reduction	Chunk latency mean/p50/max (ms)	Structure	Cache
coding-agent	5/5	—	56%	990.3/1113/1338	✅	✅
research	1/3	—	46%	1014/1014/1014	✅	✅
contradictory-updates	3/3	—	36%	469/469/469	✅	✅
tool-spam	3/3	—	70%	597.8/497/1063	✅	✅

Recall = seeded facts answered correctly from the compacted view alone. Results are comparable only within one model; see bench/README.md.