Long-context downstream performance on LongBench v1
27.5ScoreThriftAttention
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ThriftAttentionModel=Llama3.1 (8B), Budget (k)=10%2026.05 | 27.5 | 111.8 | |
| ThriftAttentionModel=Llama3.1 (8B), Budget (k)=5%2026.05 | 27.1 | 108.7 | |
| ThriftAttentionModel=Llama3.1 (8B), Budget (k)=25%2026.05 | 27 | 107.9 | |
| FP16Model=Llama3.2 (3B)2026.05 | 26.7 | 100 | |
| FP16Model=Ministral 3 (8B)2026.05 | 26.7 | 100 | |
| ThriftAttentionModel=Ministral 3 (8B), Budget (k)=25%2026.05 | 26.4 | 97.8 | |
| ThriftAttentionModel=Ministral 3 (8B), Budget (k)=10%2026.05 | 26.2 | 96.4 | |
| FP16Model=Llama3.1 (8B)2026.05 | 26 | 100 | |
| FP16Model=Qwen3 (4B)2026.05 | 26 | 100 | |
| ThriftAttentionModel=Ministral 3 (8B), Budget (k)=5%2026.05 | 25.8 | 93.4 | |
| ThriftAttentionModel=Qwen3 (4B), Budget (k)=25%2026.05 | 25.7 | 97.6 | |
| ThriftAttentionModel=Qwen3 (4B), Budget (k)=10%2026.05 | 25.6 | 96.8 | |
| ThriftAttentionModel=Qwen3 (4B), Budget (k)=5%2026.05 | 25.4 | 95.2 | |
| ThriftAttentionModel=Llama3.2 (3B), Budget (k)=5%2026.05 | 25.2 | 88.9 | |
| ThriftAttentionModel=Llama3.2 (3B), Budget (k)=10%2026.05 | 24.9 | 86.5 | |
| FP16Model=Qwen3 (8B)2026.05 | 24.6 | 100 | |
| ThriftAttentionModel=Qwen3 (8B), Budget (k)=25%2026.05 | 24.2 | 96.6 | |
| ThriftAttentionModel=Qwen3 (8B), Budget (k)=10%2026.05 | 24.1 | 95.8 | |
| ThriftAttentionModel=Llama3.2 (3B), Budget (k)=25%2026.05 | 24 | 80 | |
| ThriftAttentionModel=Qwen3 (8B), Budget (k)=5%2026.05 | 23.7 | 92.4 | |
| FP4Model=Qwen3 (4B)2026.05 | 13.4 | 0 | |
| FP4Model=Llama3.1 (8B)2026.05 | 13.3 | 0 | |
| FP4Model=Llama3.2 (3B)2026.05 | 13.2 | 0 | |
| FP4Model=Ministral 3 (8B)2026.05 | 13 | 0 | |
| FP4Model=Qwen3 (8B)2026.05 | 12.8 | 0 |