105 lines
2.3 KiB
YAML
105 lines
2.3 KiB
YAML
- model: claude-3-5-haiku
|
|
core_memory: 83.5
|
|
archival_memory: 96.33
|
|
average: 87.78
|
|
- model: gemini-2-5-pro
|
|
core_memory: 99.33
|
|
archival_memory: 96.0
|
|
average: 98.22
|
|
- model: claude-3-7-sonnet-extended
|
|
core_memory: 97.0
|
|
archival_memory: 93.33
|
|
average: 95.78
|
|
- model: gemini-2-5-flash
|
|
core_memory: 94.5
|
|
archival_memory: 93.0
|
|
average: 94.0
|
|
- model: openai-gpt-4.1
|
|
core_memory: 98.33
|
|
archival_memory: 89.67
|
|
average: 95.44
|
|
- model: claude-3-7-sonnet
|
|
core_memory: 94.83
|
|
archival_memory: 88.0
|
|
average: 92.56
|
|
- model: together-llama-4-scout-17b
|
|
core_memory: 74.67
|
|
archival_memory: 86.33
|
|
average: 78.56
|
|
- model: together-qwen-2-5-72b
|
|
core_memory: 76.5
|
|
archival_memory: 79.33
|
|
average: 77.44
|
|
- model: claude-3-5-sonnet
|
|
core_memory: 96.67
|
|
archival_memory: 76.67
|
|
average: 90.0
|
|
- model: openai-gpt-4o
|
|
core_memory: 97.5
|
|
archival_memory: 69.0
|
|
average: 88.0
|
|
- model: together-llama-3-1-405b
|
|
core_memory: 92.17
|
|
archival_memory: 60.67
|
|
average: 81.67
|
|
- model: together-llama-4-maverick-17b
|
|
core_memory: 67.0
|
|
archival_memory: 53.0
|
|
average: 62.33
|
|
- model: openai-o1
|
|
core_memory: 89.5
|
|
archival_memory: 52.33
|
|
average: 77.11
|
|
- model: openai-gpt-4.1-mini
|
|
core_memory: 96.83
|
|
archival_memory: 41.0
|
|
average: 78.22
|
|
- model: together-deepseek-v3
|
|
core_memory: 96.83
|
|
archival_memory: 26.33
|
|
average: 73.33
|
|
- model: together-llama-3-2-3b
|
|
core_memory: 0.0
|
|
archival_memory: 14.0
|
|
average: 4.67
|
|
- model: together-llama-3-70b
|
|
core_memory: 47.33
|
|
archival_memory: 13.0
|
|
average: 35.89
|
|
- model: together-meta-llama-3-1-8b
|
|
core_memory: 45.0
|
|
archival_memory: 8.0
|
|
average: 32.67
|
|
- model: together-llama-3-3-70b
|
|
core_memory: 96.33
|
|
archival_memory: 6.33
|
|
average: 66.33
|
|
- model: together-meta-llama-3-1-70b
|
|
core_memory: 90.83
|
|
archival_memory: 6.0
|
|
average: 62.56
|
|
- model: openai-o3-mini
|
|
core_memory: 95.83
|
|
archival_memory: 5.33
|
|
average: 65.67
|
|
- model: openai-o4-mini
|
|
core_memory: 98.17
|
|
archival_memory: 4.67
|
|
average: 67.0
|
|
- model: openai-gpt-4.1-nano
|
|
core_memory: 35.0
|
|
archival_memory: 2.0
|
|
average: 24.0
|
|
- model: openai-gpt-4o-mini
|
|
core_memory: 97.17
|
|
archival_memory: 1.33
|
|
average: 65.22
|
|
- model: together-qwen-2-5-7b
|
|
core_memory: 24.5
|
|
archival_memory: 1.0
|
|
average: 16.67
|
|
- model: openai-gpt-3.5-turbo
|
|
core_memory: 31.17
|
|
archival_memory: 0.67
|
|
average: 21.0
|