Reasoning on AIME 32K generation 25
74AccuracyOSCAR
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| OSCARBackbone=Qwen3-32B, Generation Length=32K, KV-cache Mode=INT22026.05 | 74 | 2.38 | |
| Original BF16Backbone=Qwen3-32B, Generation Length=32K, KV-cache Mode=BF162026.05 | 72.59 | 16 | |
| KittyBackbone=Qwen3-32B, Generation Length=32K, KV-cache Mode=INT22026.05 | 69.26 | 2.39 | |
| OSCARBackbone=Qwen3-8B, Generation Length=32K, KV-cache Mode=INT22026.05 | 66.67 | 2.38 | |
| Original BF16Backbone=Qwen3-8B, Generation Length=32K, KV-cache Mode=BF162026.05 | 66 | 16 | |
| KittyBackbone=Qwen3-8B, Generation Length=32K, KV-cache Mode=INT22026.05 | 59.67 | 2.39 | |
| KIVI-KV2*Backbone=Qwen3-32B, Generation Length=32K, KV-cache Mode=INT2 (Modified)2026.05 | 59.05 | 2.26 | |
| KIVI-KV2*Backbone=Qwen3-8B, Generation Length=32K, KV-cache Mode=INT2 (Modified)2026.05 | 57.67 | 2.26 | |
| KIVI-KV2Backbone=Qwen3-32B, Generation Length=32K, KV-cache Mode=INT22026.05 | 57.41 | 2.25 | |
| KIVI-KV2Backbone=Qwen3-8B, Generation Length=32K, KV-cache Mode=INT22026.05 | 52.33 | 2.25 |